diff --git "a/job_new.json" "b/job_new.json" --- "a/job_new.json" +++ "b/job_new.json" @@ -16,12 +16,12 @@ "progress": "finished", "status_output": false, "cal_filename": "Qwen/Qwen2.5-Coder-1.5B-Instruct-EXL2/cal_data.safetensors", - "last_module_idx": 50, + "last_module_idx": 58, "measurement": { "model.layers.0.self_attn": [ { - "accuracy": 0.9079901054501534, - "total_bits": 4051584, + "accuracy": 0.7532381117343903, + "total_bits": 12046848, "q_proj": { "group_size": { "3": 64, @@ -84,8 +84,8 @@ } }, { - "accuracy": 0.9210989624261856, - "total_bits": 4199040, + "accuracy": 0.771048441529274, + "total_bits": 12325376, "q_proj": { "group_size": { "3": 64, @@ -148,8 +148,8 @@ } }, { - "accuracy": 0.9341454803943634, - "total_bits": 4293424, + "accuracy": 0.8117339909076691, + "total_bits": 12672960, "q_proj": { "group_size": { "3": 64, @@ -212,8 +212,8 @@ } }, { - "accuracy": 0.9540616944432259, - "total_bits": 5070816, + "accuracy": 0.8756691366434097, + "total_bits": 15056256, "q_proj": { "group_size": { "3": 64, @@ -276,8 +276,8 @@ } }, { - "accuracy": 0.9560479000210762, - "total_bits": 5958976, + "accuracy": 0.8801903575658798, + "total_bits": 17836288, "q_proj": { "group_size": { "4": 128, @@ -340,8 +340,8 @@ } }, { - "accuracy": 0.9575468599796295, - "total_bits": 5962896, + "accuracy": 0.8807467222213745, + "total_bits": 17832768, "q_proj": { "group_size": { "4": 128, @@ -404,8 +404,8 @@ } }, { - "accuracy": 0.9709379449486732, - "total_bits": 7531840, + "accuracy": 0.9321320429444313, + "total_bits": 22423808, "q_proj": { "group_size": { "4": 128 @@ -456,8 +456,8 @@ } }, { - "accuracy": 0.9723670464009047, - "total_bits": 7535760, + "accuracy": 0.9347590953111649, + "total_bits": 22436672, "q_proj": { "group_size": { "4": 128 @@ -508,8 +508,8 @@ } }, { - "accuracy": 0.9751146323978901, - "total_bits": 7598368, + "accuracy": 0.9393426515161991, + "total_bits": 22623872, "q_proj": { "group_size": { "4": 64 @@ -560,8 +560,8 @@ } }, { - "accuracy": 0.9763321615755558, - "total_bits": 7707904, + "accuracy": 0.9416504055261612, + "total_bits": 22946816, "q_proj": { "group_size": { "4": 32 @@ -612,8 +612,8 @@ } }, { - "accuracy": 0.9784583374857903, - "total_bits": 7797904, + "accuracy": 0.9402352832257748, + "total_bits": 23337792, "q_proj": { "group_size": { "5": 128, @@ -676,8 +676,8 @@ } }, { - "accuracy": 0.9799316953867674, - "total_bits": 7856416, + "accuracy": 0.9436496049165726, + "total_bits": 23303808, "q_proj": { "group_size": { "5": 64, @@ -740,8 +740,8 @@ } }, { - "accuracy": 0.9823341108858585, - "total_bits": 7950976, + "accuracy": 0.954001359641552, + "total_bits": 23630336, "q_proj": { "group_size": { "5": 64, @@ -801,8 +801,8 @@ } }, { - "accuracy": 0.9834211450070143, - "total_bits": 8006912, + "accuracy": 0.9576069973409176, + "total_bits": 23872512, "q_proj": { "group_size": { "5": 32, @@ -862,8 +862,8 @@ } }, { - "accuracy": 0.9901147726923227, - "total_bits": 9727296, + "accuracy": 0.9751354176551104, + "total_bits": 29174016, "q_proj": { "group_size": { "6": 128, @@ -923,8 +923,8 @@ } }, { - "accuracy": 0.9916111528873444, - "total_bits": 9841920, + "accuracy": 0.9783839359879494, + "total_bits": 29377536, "q_proj": { "group_size": { "6": 32, @@ -984,8 +984,8 @@ } }, { - "accuracy": 0.9924060190096498, - "total_bits": 11201856, + "accuracy": 0.982197217643261, + "total_bits": 33433856, "q_proj": { "group_size": { "6": 128 @@ -1036,8 +1036,8 @@ } }, { - "accuracy": 0.9956152755767107, - "total_bits": 11607296, + "accuracy": 0.9896592563018203, + "total_bits": 34743296, "q_proj": { "group_size": { "6": 32 @@ -1088,8 +1088,8 @@ } }, { - "accuracy": 0.9979100672062486, - "total_bits": 14871872, + "accuracy": 0.9954520598985255, + "total_bits": 44443904, "q_proj": { "group_size": { "8": 128 @@ -1142,8 +1142,8 @@ ], "model.layers.0.mlp": [ { - "accuracy": 0.7537962645292282, - "total_bits": 29656128, + "accuracy": 0.8799141943454742, + "total_bits": 93392016, "gate_proj": { "group_size": { "3": 64, @@ -1194,8 +1194,8 @@ } }, { - "accuracy": 0.7683756947517395, - "total_bits": 30901312, + "accuracy": 0.8825936689972878, + "total_bits": 96259216, "gate_proj": { "group_size": { "3": 64, @@ -1246,8 +1246,8 @@ } }, { - "accuracy": 0.7968224436044693, - "total_bits": 34456000, + "accuracy": 0.9097775220870972, + "total_bits": 107863424, "gate_proj": { "group_size": { "3": 64, @@ -1295,8 +1295,8 @@ } }, { - "accuracy": 0.8044690489768982, - "total_bits": 38584768, + "accuracy": 0.9192143902182579, + "total_bits": 120937856, "gate_proj": { "group_size": { "3": 64, @@ -1344,8 +1344,8 @@ } }, { - "accuracy": 0.8812098801136017, - "total_bits": 43342560, + "accuracy": 0.9402356259524822, + "total_bits": 136162224, "gate_proj": { "group_size": { "4": 128, @@ -1396,8 +1396,8 @@ } }, { - "accuracy": 0.8911077007651329, - "total_bits": 44242688, + "accuracy": 0.945199016481638, + "total_bits": 139118208, "gate_proj": { "group_size": { "4": 32, @@ -1448,8 +1448,8 @@ } }, { - "accuracy": 0.9013418033719063, - "total_bits": 47520448, + "accuracy": 0.957384280860424, + "total_bits": 149584176, "gate_proj": { "group_size": { "4": 32, @@ -1497,8 +1497,8 @@ } }, { - "accuracy": 0.9381188377737999, - "total_bits": 54270672, + "accuracy": 0.9680311493575573, + "total_bits": 170943088, "gate_proj": { "group_size": { "4": 128 @@ -1540,8 +1540,8 @@ } }, { - "accuracy": 0.9429273754358292, - "total_bits": 55072512, + "accuracy": 0.9716346263885498, + "total_bits": 173467264, "gate_proj": { "group_size": { "4": 32 @@ -1583,8 +1583,8 @@ } }, { - "accuracy": 0.9395514503121376, - "total_bits": 56187616, + "accuracy": 0.9696477428078651, + "total_bits": 176761776, "gate_proj": { "group_size": { "5": 128, @@ -1635,8 +1635,8 @@ } }, { - "accuracy": 0.9473779834806919, - "total_bits": 57087744, + "accuracy": 0.9736586231738329, + "total_bits": 179717760, "gate_proj": { "group_size": { "5": 32, @@ -1687,8 +1687,8 @@ } }, { - "accuracy": 0.9691675771027803, - "total_bits": 69032672, + "accuracy": 0.9845514940097928, + "total_bits": 217361328, "gate_proj": { "group_size": { "6": 128, @@ -1739,8 +1739,8 @@ } }, { - "accuracy": 0.9737564735114574, - "total_bits": 69932800, + "accuracy": 0.9868995463475585, + "total_bits": 220317312, "gate_proj": { "group_size": { "6": 32, @@ -1791,8 +1791,8 @@ } }, { - "accuracy": 0.9826037380844355, - "total_bits": 79551200, + "accuracy": 0.9915019096806645, + "total_bits": 250850224, "gate_proj": { "group_size": { "6": 128 @@ -1834,8 +1834,8 @@ } }, { - "accuracy": 0.9841426350176334, - "total_bits": 82937280, + "accuracy": 0.9922892493195832, + "total_bits": 262906272, "gate_proj": { "group_size": { "8": 128, @@ -1883,8 +1883,8 @@ } }, { - "accuracy": 0.986230812035501, - "total_bits": 90277312, + "accuracy": 0.9949854775331914, + "total_bits": 286106016, "gate_proj": { "group_size": { "8": 128, @@ -1929,8 +1929,8 @@ } }, { - "accuracy": 0.9954390907660127, - "total_bits": 105219520, + "accuracy": 0.997696247883141, + "total_bits": 331981216, "gate_proj": { "group_size": { "8": 128 @@ -1971,8 +1971,8 @@ ], "model.layers.1.self_attn": [ { - "accuracy": 0.9195414483547211, - "total_bits": 4051584, + "accuracy": 0.9333188757300377, + "total_bits": 12046848, "q_proj": { "group_size": { "3": 64, @@ -2035,8 +2035,8 @@ } }, { - "accuracy": 0.927167534828186, - "total_bits": 4199040, + "accuracy": 0.9370334222912788, + "total_bits": 12325376, "q_proj": { "group_size": { "3": 64, @@ -2099,8 +2099,8 @@ } }, { - "accuracy": 0.9342408627271652, - "total_bits": 4293424, + "accuracy": 0.9441889226436615, + "total_bits": 12672960, "q_proj": { "group_size": { "3": 64, @@ -2163,8 +2163,8 @@ } }, { - "accuracy": 0.9536051601171494, - "total_bits": 5070816, + "accuracy": 0.9656500816345215, + "total_bits": 15056256, "q_proj": { "group_size": { "3": 64, @@ -2227,8 +2227,8 @@ } }, { - "accuracy": 0.9601038135588169, - "total_bits": 5958976, + "accuracy": 0.9683590605854988, + "total_bits": 17836288, "q_proj": { "group_size": { "4": 128, @@ -2291,8 +2291,8 @@ } }, { - "accuracy": 0.9608978591859341, - "total_bits": 5962896, + "accuracy": 0.9685361199080944, + "total_bits": 17832768, "q_proj": { "group_size": { "4": 128, @@ -2355,8 +2355,8 @@ } }, { - "accuracy": 0.9768611993640661, - "total_bits": 7531840, + "accuracy": 0.981427364051342, + "total_bits": 22423808, "q_proj": { "group_size": { "4": 128 @@ -2407,8 +2407,8 @@ } }, { - "accuracy": 0.9775344431400299, - "total_bits": 7535760, + "accuracy": 0.9818543680012226, + "total_bits": 22436672, "q_proj": { "group_size": { "4": 128 @@ -2459,8 +2459,8 @@ } }, { - "accuracy": 0.9794217944145203, - "total_bits": 7598368, + "accuracy": 0.9830422420054674, + "total_bits": 22623872, "q_proj": { "group_size": { "4": 64 @@ -2511,8 +2511,8 @@ } }, { - "accuracy": 0.9805951975286007, - "total_bits": 7707904, + "accuracy": 0.9840035960078239, + "total_bits": 22946816, "q_proj": { "group_size": { "4": 32 @@ -2563,8 +2563,8 @@ } }, { - "accuracy": 0.9801530670374632, - "total_bits": 7797904, + "accuracy": 0.9840645231306553, + "total_bits": 23337792, "q_proj": { "group_size": { "5": 128, @@ -2627,8 +2627,8 @@ } }, { - "accuracy": 0.9814895000308752, - "total_bits": 7856416, + "accuracy": 0.9845752185210586, + "total_bits": 23303808, "q_proj": { "group_size": { "5": 64, @@ -2691,8 +2691,8 @@ } }, { - "accuracy": 0.9834644962102175, - "total_bits": 7950976, + "accuracy": 0.9861648315563798, + "total_bits": 23630336, "q_proj": { "group_size": { "5": 64, @@ -2752,8 +2752,8 @@ } }, { - "accuracy": 0.9845743589103222, - "total_bits": 8006912, + "accuracy": 0.9870333140715957, + "total_bits": 23872512, "q_proj": { "group_size": { "5": 32, @@ -2813,8 +2813,8 @@ } }, { - "accuracy": 0.9909268477931619, - "total_bits": 9727296, + "accuracy": 0.9927375833503902, + "total_bits": 29174016, "q_proj": { "group_size": { "6": 128, @@ -2874,8 +2874,8 @@ } }, { - "accuracy": 0.9922212068922818, - "total_bits": 9841920, + "accuracy": 0.9935035100206733, + "total_bits": 29377536, "q_proj": { "group_size": { "6": 32, @@ -2935,8 +2935,8 @@ } }, { - "accuracy": 0.9939816757105291, - "total_bits": 11201856, + "accuracy": 0.9951658211648464, + "total_bits": 33433856, "q_proj": { "group_size": { "6": 128 @@ -2987,8 +2987,8 @@ } }, { - "accuracy": 0.9959470983594656, - "total_bits": 11607296, + "accuracy": 0.9966006258036941, + "total_bits": 34743296, "q_proj": { "group_size": { "6": 32 @@ -3039,8 +3039,8 @@ } }, { - "accuracy": 0.9983556552324444, - "total_bits": 14871872, + "accuracy": 0.9986925142584369, + "total_bits": 44443904, "q_proj": { "group_size": { "8": 128 @@ -3093,8 +3093,8 @@ ], "model.layers.1.mlp": [ { - "accuracy": 0.8358551561832428, - "total_bits": 29656128, + "accuracy": 0.8831949979066849, + "total_bits": 93392016, "gate_proj": { "group_size": { "3": 64, @@ -3145,8 +3145,8 @@ } }, { - "accuracy": 0.8424440324306488, - "total_bits": 30901312, + "accuracy": 0.8892563432455063, + "total_bits": 96259216, "gate_proj": { "group_size": { "3": 64, @@ -3197,8 +3197,8 @@ } }, { - "accuracy": 0.862356960773468, - "total_bits": 34456000, + "accuracy": 0.8948070108890533, + "total_bits": 107863424, "gate_proj": { "group_size": { "3": 64, @@ -3246,8 +3246,8 @@ } }, { - "accuracy": 0.8681714236736298, - "total_bits": 38584768, + "accuracy": 0.8964388519525528, + "total_bits": 120937856, "gate_proj": { "group_size": { "3": 64, @@ -3295,8 +3295,8 @@ } }, { - "accuracy": 0.9167918860912323, - "total_bits": 43342560, + "accuracy": 0.9603819288313389, + "total_bits": 136162224, "gate_proj": { "group_size": { "4": 128, @@ -3347,8 +3347,8 @@ } }, { - "accuracy": 0.9238021150231361, - "total_bits": 44242688, + "accuracy": 0.9561992287635803, + "total_bits": 139118208, "gate_proj": { "group_size": { "4": 32, @@ -3399,8 +3399,8 @@ } }, { - "accuracy": 0.9318408221006393, - "total_bits": 47520448, + "accuracy": 0.9582490734755993, + "total_bits": 149584176, "gate_proj": { "group_size": { "4": 32, @@ -3448,8 +3448,8 @@ } }, { - "accuracy": 0.9563885070383549, - "total_bits": 54270672, + "accuracy": 0.9812938813120127, + "total_bits": 170943088, "gate_proj": { "group_size": { "4": 128 @@ -3491,8 +3491,8 @@ } }, { - "accuracy": 0.9603695422410965, - "total_bits": 55072512, + "accuracy": 0.9806262496858835, + "total_bits": 173467264, "gate_proj": { "group_size": { "4": 32 @@ -3534,8 +3534,8 @@ } }, { - "accuracy": 0.9573803171515465, - "total_bits": 56187616, + "accuracy": 0.9827004242688417, + "total_bits": 176761776, "gate_proj": { "group_size": { "5": 128, @@ -3586,8 +3586,8 @@ } }, { - "accuracy": 0.9629036076366901, - "total_bits": 57087744, + "accuracy": 0.9846479566767812, + "total_bits": 179717760, "gate_proj": { "group_size": { "5": 32, @@ -3638,8 +3638,8 @@ } }, { - "accuracy": 0.978231294080615, - "total_bits": 69032672, + "accuracy": 0.9912021830677986, + "total_bits": 217361328, "gate_proj": { "group_size": { "6": 128, @@ -3690,8 +3690,8 @@ } }, { - "accuracy": 0.9814751539379358, - "total_bits": 69932800, + "accuracy": 0.9927156791090965, + "total_bits": 220317312, "gate_proj": { "group_size": { "6": 32, @@ -3742,8 +3742,8 @@ } }, { - "accuracy": 0.9879805697128177, - "total_bits": 79551200, + "accuracy": 0.9945937651209533, + "total_bits": 250850224, "gate_proj": { "group_size": { "6": 128 @@ -3785,8 +3785,8 @@ } }, { - "accuracy": 0.9888631301000714, - "total_bits": 82937280, + "accuracy": 0.9958127764984965, + "total_bits": 262906272, "gate_proj": { "group_size": { "8": 128, @@ -3834,8 +3834,8 @@ } }, { - "accuracy": 0.9905711691826582, - "total_bits": 90277312, + "accuracy": 0.9965894345659763, + "total_bits": 286106016, "gate_proj": { "group_size": { "8": 128, @@ -3880,8 +3880,8 @@ } }, { - "accuracy": 0.99683584459126, - "total_bits": 105219520, + "accuracy": 0.9983784968499094, + "total_bits": 331981216, "gate_proj": { "group_size": { "8": 128 @@ -3922,8 +3922,8 @@ ], "model.layers.2.self_attn": [ { - "accuracy": 0.9125006794929504, - "total_bits": 4051584, + "accuracy": 0.9589957632124424, + "total_bits": 12046848, "q_proj": { "group_size": { "3": 64, @@ -3986,8 +3986,8 @@ } }, { - "accuracy": 0.9216000959277153, - "total_bits": 4199040, + "accuracy": 0.9602763131260872, + "total_bits": 12325376, "q_proj": { "group_size": { "3": 64, @@ -4050,8 +4050,8 @@ } }, { - "accuracy": 0.9291190356016159, - "total_bits": 4293424, + "accuracy": 0.9645074233412743, + "total_bits": 12672960, "q_proj": { "group_size": { "3": 64, @@ -4114,8 +4114,8 @@ } }, { - "accuracy": 0.9468126408755779, - "total_bits": 5070816, + "accuracy": 0.97387027554214, + "total_bits": 15056256, "q_proj": { "group_size": { "3": 64, @@ -4178,8 +4178,8 @@ } }, { - "accuracy": 0.9568752348423004, - "total_bits": 5958976, + "accuracy": 0.9791378080844879, + "total_bits": 17836288, "q_proj": { "group_size": { "4": 128, @@ -4242,8 +4242,8 @@ } }, { - "accuracy": 0.9577730931341648, - "total_bits": 5962896, + "accuracy": 0.9796560574322939, + "total_bits": 17832768, "q_proj": { "group_size": { "4": 128, @@ -4306,8 +4306,8 @@ } }, { - "accuracy": 0.9726411122828722, - "total_bits": 7531840, + "accuracy": 0.9867294300347567, + "total_bits": 22423808, "q_proj": { "group_size": { "4": 128 @@ -4358,8 +4358,8 @@ } }, { - "accuracy": 0.9735233001410961, - "total_bits": 7535760, + "accuracy": 0.9874424943700433, + "total_bits": 22436672, "q_proj": { "group_size": { "4": 128 @@ -4410,8 +4410,8 @@ } }, { - "accuracy": 0.9760294500738382, - "total_bits": 7598368, + "accuracy": 0.9883597102016211, + "total_bits": 22623872, "q_proj": { "group_size": { "4": 64 @@ -4462,8 +4462,8 @@ } }, { - "accuracy": 0.9775469116866589, - "total_bits": 7707904, + "accuracy": 0.9882704289630055, + "total_bits": 22946816, "q_proj": { "group_size": { "4": 32 @@ -4514,8 +4514,8 @@ } }, { - "accuracy": 0.9784345105290413, - "total_bits": 7797904, + "accuracy": 0.9896981036290526, + "total_bits": 23337792, "q_proj": { "group_size": { "5": 128, @@ -4578,8 +4578,8 @@ } }, { - "accuracy": 0.9798352308571339, - "total_bits": 7856416, + "accuracy": 0.9900533556938171, + "total_bits": 23303808, "q_proj": { "group_size": { "5": 64, @@ -4642,8 +4642,8 @@ } }, { - "accuracy": 0.9814392048865557, - "total_bits": 7950976, + "accuracy": 0.990645244717598, + "total_bits": 23630336, "q_proj": { "group_size": { "5": 64, @@ -4703,8 +4703,8 @@ } }, { - "accuracy": 0.9829360991716385, - "total_bits": 8006912, + "accuracy": 0.9914778620004654, + "total_bits": 23872512, "q_proj": { "group_size": { "5": 32, @@ -4764,8 +4764,8 @@ } }, { - "accuracy": 0.9897045381367207, - "total_bits": 9727296, + "accuracy": 0.9951063548214734, + "total_bits": 29174016, "q_proj": { "group_size": { "6": 128, @@ -4825,8 +4825,8 @@ } }, { - "accuracy": 0.9913953728973866, - "total_bits": 9841920, + "accuracy": 0.9957435643300414, + "total_bits": 29377536, "q_proj": { "group_size": { "6": 32, @@ -4886,8 +4886,8 @@ } }, { - "accuracy": 0.9928306452929974, - "total_bits": 11201856, + "accuracy": 0.9966391315683722, + "total_bits": 33433856, "q_proj": { "group_size": { "6": 128 @@ -4938,8 +4938,8 @@ } }, { - "accuracy": 0.9953591991215944, - "total_bits": 11607296, + "accuracy": 0.997647543437779, + "total_bits": 34743296, "q_proj": { "group_size": { "6": 32 @@ -4990,8 +4990,8 @@ } }, { - "accuracy": 0.998056368320249, - "total_bits": 14871872, + "accuracy": 0.9991065739304759, + "total_bits": 44443904, "q_proj": { "group_size": { "8": 128 @@ -5044,8 +5044,8 @@ ], "model.layers.2.mlp": [ { - "accuracy": 0.810150995850563, - "total_bits": 29656128, + "accuracy": 0.9670937694609165, + "total_bits": 93392016, "gate_proj": { "group_size": { "3": 64, @@ -5096,8 +5096,8 @@ } }, { - "accuracy": 0.8221035897731781, - "total_bits": 30901312, + "accuracy": 0.9681853204965591, + "total_bits": 96259216, "gate_proj": { "group_size": { "3": 64, @@ -5148,8 +5148,8 @@ } }, { - "accuracy": 0.8363298326730728, - "total_bits": 34456000, + "accuracy": 0.9733225479722023, + "total_bits": 107863424, "gate_proj": { "group_size": { "3": 64, @@ -5197,8 +5197,8 @@ } }, { - "accuracy": 0.8381976932287216, - "total_bits": 38584768, + "accuracy": 0.9762210380285978, + "total_bits": 120937856, "gate_proj": { "group_size": { "3": 64, @@ -5246,8 +5246,8 @@ } }, { - "accuracy": 0.9282058924436569, - "total_bits": 43342560, + "accuracy": 0.9817080665379763, + "total_bits": 136162224, "gate_proj": { "group_size": { "4": 128, @@ -5298,8 +5298,8 @@ } }, { - "accuracy": 0.9362002462148666, - "total_bits": 44242688, + "accuracy": 0.9838489722460508, + "total_bits": 139118208, "gate_proj": { "group_size": { "4": 32, @@ -5350,8 +5350,8 @@ } }, { - "accuracy": 0.9399041309952736, - "total_bits": 47520448, + "accuracy": 0.9864885183051229, + "total_bits": 149584176, "gate_proj": { "group_size": { "4": 32, @@ -5399,8 +5399,8 @@ } }, { - "accuracy": 0.9605619125068188, - "total_bits": 54270672, + "accuracy": 0.9895894071087241, + "total_bits": 170943088, "gate_proj": { "group_size": { "4": 128 @@ -5442,8 +5442,8 @@ } }, { - "accuracy": 0.9644056335091591, - "total_bits": 55072512, + "accuracy": 0.9904790800064802, + "total_bits": 173467264, "gate_proj": { "group_size": { "4": 32 @@ -5485,8 +5485,8 @@ } }, { - "accuracy": 0.9707082081586123, - "total_bits": 56187616, + "accuracy": 0.9899621969088912, + "total_bits": 176761776, "gate_proj": { "group_size": { "5": 128, @@ -5537,8 +5537,8 @@ } }, { - "accuracy": 0.9749497789889574, - "total_bits": 57087744, + "accuracy": 0.9908143579959869, + "total_bits": 179717760, "gate_proj": { "group_size": { "5": 32, @@ -5589,8 +5589,8 @@ } }, { - "accuracy": 0.9782808125019073, - "total_bits": 69032672, + "accuracy": 0.995028751436621, + "total_bits": 217361328, "gate_proj": { "group_size": { "6": 128, @@ -5641,8 +5641,8 @@ } }, { - "accuracy": 0.9880842100828886, - "total_bits": 69932800, + "accuracy": 0.9951841593720019, + "total_bits": 220317312, "gate_proj": { "group_size": { "6": 32, @@ -5693,8 +5693,8 @@ } }, { - "accuracy": 0.9875405784696341, - "total_bits": 79551200, + "accuracy": 0.9965546226594597, + "total_bits": 250850224, "gate_proj": { "group_size": { "6": 128 @@ -5736,8 +5736,8 @@ } }, { - "accuracy": 0.9918466089293361, - "total_bits": 82937280, + "accuracy": 0.9976431999821216, + "total_bits": 262906272, "gate_proj": { "group_size": { "8": 128, @@ -5785,8 +5785,8 @@ } }, { - "accuracy": 0.9928205446340144, - "total_bits": 90277312, + "accuracy": 0.9983571015764028, + "total_bits": 286106016, "gate_proj": { "group_size": { "8": 128, @@ -5831,8 +5831,8 @@ } }, { - "accuracy": 0.9975450397469103, - "total_bits": 105219520, + "accuracy": 0.9991033712285571, + "total_bits": 331981216, "gate_proj": { "group_size": { "8": 128 @@ -5873,8 +5873,8 @@ ], "model.layers.3.self_attn": [ { - "accuracy": 0.9153547659516335, - "total_bits": 4051584, + "accuracy": 0.9806258715689182, + "total_bits": 12046848, "q_proj": { "group_size": { "3": 64, @@ -5937,8 +5937,8 @@ } }, { - "accuracy": 0.9326226338744164, - "total_bits": 4199040, + "accuracy": 0.9815356396138668, + "total_bits": 12325376, "q_proj": { "group_size": { "3": 64, @@ -6001,8 +6001,8 @@ } }, { - "accuracy": 0.9325419664382935, - "total_bits": 4293424, + "accuracy": 0.9832805767655373, + "total_bits": 12672960, "q_proj": { "group_size": { "3": 64, @@ -6065,8 +6065,8 @@ } }, { - "accuracy": 0.93686393648386, - "total_bits": 5070816, + "accuracy": 0.987049943767488, + "total_bits": 15056256, "q_proj": { "group_size": { "3": 64, @@ -6129,8 +6129,8 @@ } }, { - "accuracy": 0.954244676977396, - "total_bits": 5958976, + "accuracy": 0.9905584482476115, + "total_bits": 17836288, "q_proj": { "group_size": { "4": 128, @@ -6193,8 +6193,8 @@ } }, { - "accuracy": 0.9559049680829048, - "total_bits": 5962896, + "accuracy": 0.990821236744523, + "total_bits": 17832768, "q_proj": { "group_size": { "4": 128, @@ -6257,8 +6257,8 @@ } }, { - "accuracy": 0.9760886766016483, - "total_bits": 7531840, + "accuracy": 0.9935666471719742, + "total_bits": 22423808, "q_proj": { "group_size": { "4": 128 @@ -6309,8 +6309,8 @@ } }, { - "accuracy": 0.9757480621337891, - "total_bits": 7535760, + "accuracy": 0.9939650213345885, + "total_bits": 22436672, "q_proj": { "group_size": { "4": 128 @@ -6361,8 +6361,8 @@ } }, { - "accuracy": 0.9783012196421623, - "total_bits": 7598368, + "accuracy": 0.9940219195559621, + "total_bits": 22623872, "q_proj": { "group_size": { "4": 64 @@ -6413,8 +6413,8 @@ } }, { - "accuracy": 0.9781928192824125, - "total_bits": 7707904, + "accuracy": 0.9947303617373109, + "total_bits": 22946816, "q_proj": { "group_size": { "4": 32 @@ -6465,8 +6465,8 @@ } }, { - "accuracy": 0.9772706441581249, - "total_bits": 7797904, + "accuracy": 0.9952523000538349, + "total_bits": 23337792, "q_proj": { "group_size": { "5": 128, @@ -6529,8 +6529,8 @@ } }, { - "accuracy": 0.9772896263748407, - "total_bits": 7856416, + "accuracy": 0.9955991068854928, + "total_bits": 23303808, "q_proj": { "group_size": { "5": 64, @@ -6593,8 +6593,8 @@ } }, { - "accuracy": 0.9815661571919918, - "total_bits": 7950976, + "accuracy": 0.9957800186239183, + "total_bits": 23630336, "q_proj": { "group_size": { "5": 64, @@ -6654,8 +6654,8 @@ } }, { - "accuracy": 0.9867081800475717, - "total_bits": 8006912, + "accuracy": 0.9963479884900153, + "total_bits": 23872512, "q_proj": { "group_size": { "5": 32, @@ -6715,8 +6715,8 @@ } }, { - "accuracy": 0.991348922252655, - "total_bits": 9727296, + "accuracy": 0.9976901353802532, + "total_bits": 29174016, "q_proj": { "group_size": { "6": 128, @@ -6776,8 +6776,8 @@ } }, { - "accuracy": 0.9936502035707235, - "total_bits": 9841920, + "accuracy": 0.9981805344577879, + "total_bits": 29377536, "q_proj": { "group_size": { "6": 32, @@ -6837,8 +6837,8 @@ } }, { - "accuracy": 0.993813291657716, - "total_bits": 11201856, + "accuracy": 0.9982859269948676, + "total_bits": 33433856, "q_proj": { "group_size": { "6": 128 @@ -6889,8 +6889,8 @@ } }, { - "accuracy": 0.9968086224980652, - "total_bits": 11607296, + "accuracy": 0.9990541480947286, + "total_bits": 34743296, "q_proj": { "group_size": { "6": 32 @@ -6941,8 +6941,8 @@ } }, { - "accuracy": 0.998419345007278, - "total_bits": 14871872, + "accuracy": 0.9995221581193618, + "total_bits": 44443904, "q_proj": { "group_size": { "8": 128 @@ -6995,8 +6995,8 @@ ], "model.layers.3.mlp": [ { - "accuracy": 0.9348810240626335, - "total_bits": 29656128, + "accuracy": 0.967471681535244, + "total_bits": 93392016, "gate_proj": { "group_size": { "3": 64, @@ -7047,8 +7047,8 @@ } }, { - "accuracy": 0.9375668726861477, - "total_bits": 30901312, + "accuracy": 0.9686100594699383, + "total_bits": 96259216, "gate_proj": { "group_size": { "3": 64, @@ -7099,8 +7099,8 @@ } }, { - "accuracy": 0.9402441941201687, - "total_bits": 34456000, + "accuracy": 0.9723133072257042, + "total_bits": 107863424, "gate_proj": { "group_size": { "3": 64, @@ -7148,8 +7148,8 @@ } }, { - "accuracy": 0.9420037120580673, - "total_bits": 38584768, + "accuracy": 0.9736907985061407, + "total_bits": 120937856, "gate_proj": { "group_size": { "3": 64, @@ -7197,8 +7197,8 @@ } }, { - "accuracy": 0.9592516869306564, - "total_bits": 43342560, + "accuracy": 0.9834474325180054, + "total_bits": 136162224, "gate_proj": { "group_size": { "4": 128, @@ -7249,8 +7249,8 @@ } }, { - "accuracy": 0.9732199236750603, - "total_bits": 44242688, + "accuracy": 0.9849548544734716, + "total_bits": 139118208, "gate_proj": { "group_size": { "4": 32, @@ -7301,8 +7301,8 @@ } }, { - "accuracy": 0.9763146564364433, - "total_bits": 47520448, + "accuracy": 0.9864925723522902, + "total_bits": 149584176, "gate_proj": { "group_size": { "4": 32, @@ -7350,8 +7350,8 @@ } }, { - "accuracy": 0.9847144680097699, - "total_bits": 54270672, + "accuracy": 0.9909760244190693, + "total_bits": 170943088, "gate_proj": { "group_size": { "4": 128 @@ -7393,8 +7393,8 @@ } }, { - "accuracy": 0.9852770129218698, - "total_bits": 55072512, + "accuracy": 0.9919050661846995, + "total_bits": 173467264, "gate_proj": { "group_size": { "4": 32 @@ -7436,8 +7436,8 @@ } }, { - "accuracy": 0.9735454972833395, - "total_bits": 56187616, + "accuracy": 0.9915664941072464, + "total_bits": 176761776, "gate_proj": { "group_size": { "5": 128, @@ -7488,8 +7488,8 @@ } }, { - "accuracy": 0.9853277485817671, - "total_bits": 57087744, + "accuracy": 0.9926607459783554, + "total_bits": 179717760, "gate_proj": { "group_size": { "5": 32, @@ -7540,8 +7540,8 @@ } }, { - "accuracy": 0.9912315523251891, - "total_bits": 69032672, + "accuracy": 0.9956699670292437, + "total_bits": 217361328, "gate_proj": { "group_size": { "6": 128, @@ -7592,8 +7592,8 @@ } }, { - "accuracy": 0.9935814472846687, - "total_bits": 69932800, + "accuracy": 0.9963369315955788, + "total_bits": 220317312, "gate_proj": { "group_size": { "6": 32, @@ -7644,8 +7644,8 @@ } }, { - "accuracy": 0.9937410796992481, - "total_bits": 79551200, + "accuracy": 0.9975030701607466, + "total_bits": 250850224, "gate_proj": { "group_size": { "6": 128 @@ -7687,8 +7687,8 @@ } }, { - "accuracy": 0.9957921644672751, - "total_bits": 82937280, + "accuracy": 0.9978356482461095, + "total_bits": 262906272, "gate_proj": { "group_size": { "8": 128, @@ -7736,8 +7736,8 @@ } }, { - "accuracy": 0.9964254081714898, - "total_bits": 90277312, + "accuracy": 0.9982005214551464, + "total_bits": 286106016, "gate_proj": { "group_size": { "8": 128, @@ -7782,8 +7782,8 @@ } }, { - "accuracy": 0.997894095024094, - "total_bits": 105219520, + "accuracy": 0.9992947833961807, + "total_bits": 331981216, "gate_proj": { "group_size": { "8": 128 @@ -7824,8 +7824,8 @@ ], "model.layers.4.self_attn": [ { - "accuracy": 0.965219397097826, - "total_bits": 4051584, + "accuracy": 0.9799280744045973, + "total_bits": 12046848, "q_proj": { "group_size": { "3": 64, @@ -7888,8 +7888,8 @@ } }, { - "accuracy": 0.9672110378742218, - "total_bits": 4199040, + "accuracy": 0.9808804616332054, + "total_bits": 12325376, "q_proj": { "group_size": { "3": 64, @@ -7952,8 +7952,8 @@ } }, { - "accuracy": 0.9725755732506514, - "total_bits": 4293424, + "accuracy": 0.9829100053757429, + "total_bits": 12672960, "q_proj": { "group_size": { "3": 64, @@ -8016,8 +8016,8 @@ } }, { - "accuracy": 0.9767198767513037, - "total_bits": 5070816, + "accuracy": 0.9874656787142158, + "total_bits": 15056256, "q_proj": { "group_size": { "3": 64, @@ -8080,8 +8080,8 @@ } }, { - "accuracy": 0.9809905253350735, - "total_bits": 5958976, + "accuracy": 0.9895919673144817, + "total_bits": 17836288, "q_proj": { "group_size": { "4": 128, @@ -8144,8 +8144,8 @@ } }, { - "accuracy": 0.9819034244865179, - "total_bits": 5962896, + "accuracy": 0.9896589415147901, + "total_bits": 17832768, "q_proj": { "group_size": { "4": 128, @@ -8208,8 +8208,8 @@ } }, { - "accuracy": 0.9842934813350439, - "total_bits": 7531840, + "accuracy": 0.9936666111461818, + "total_bits": 22423808, "q_proj": { "group_size": { "4": 128 @@ -8260,8 +8260,8 @@ } }, { - "accuracy": 0.9853808153420687, - "total_bits": 7535760, + "accuracy": 0.9938310543075204, + "total_bits": 22436672, "q_proj": { "group_size": { "4": 128 @@ -8312,8 +8312,8 @@ } }, { - "accuracy": 0.9861455457285047, - "total_bits": 7598368, + "accuracy": 0.9945947011001408, + "total_bits": 22623872, "q_proj": { "group_size": { "4": 64 @@ -8364,8 +8364,8 @@ } }, { - "accuracy": 0.9867356372997165, - "total_bits": 7707904, + "accuracy": 0.9949558670632541, + "total_bits": 22946816, "q_proj": { "group_size": { "4": 32 @@ -8416,8 +8416,8 @@ } }, { - "accuracy": 0.9905946915969253, - "total_bits": 7797904, + "accuracy": 0.9948694771155715, + "total_bits": 23337792, "q_proj": { "group_size": { "5": 128, @@ -8480,8 +8480,8 @@ } }, { - "accuracy": 0.9913207525387406, - "total_bits": 7856416, + "accuracy": 0.9952194043435156, + "total_bits": 23303808, "q_proj": { "group_size": { "5": 64, @@ -8544,8 +8544,8 @@ } }, { - "accuracy": 0.9917500885203481, - "total_bits": 7950976, + "accuracy": 0.995673674158752, + "total_bits": 23630336, "q_proj": { "group_size": { "5": 64, @@ -8605,8 +8605,8 @@ } }, { - "accuracy": 0.9927444634959102, - "total_bits": 8006912, + "accuracy": 0.9959461484104395, + "total_bits": 23872512, "q_proj": { "group_size": { "5": 32, @@ -8666,8 +8666,8 @@ } }, { - "accuracy": 0.9950412404723465, - "total_bits": 9727296, + "accuracy": 0.9976396642159671, + "total_bits": 29174016, "q_proj": { "group_size": { "6": 128, @@ -8727,8 +8727,8 @@ } }, { - "accuracy": 0.9962598541751504, - "total_bits": 9841920, + "accuracy": 0.9979838721919805, + "total_bits": 29377536, "q_proj": { "group_size": { "6": 32, @@ -8788,8 +8788,8 @@ } }, { - "accuracy": 0.9955207225866616, - "total_bits": 11201856, + "accuracy": 0.9984299624338746, + "total_bits": 33433856, "q_proj": { "group_size": { "6": 128 @@ -8840,8 +8840,8 @@ } }, { - "accuracy": 0.9981417878298089, - "total_bits": 11607296, + "accuracy": 0.9989312083926052, + "total_bits": 34743296, "q_proj": { "group_size": { "6": 32 @@ -8892,8 +8892,8 @@ } }, { - "accuracy": 0.9989996349904686, - "total_bits": 14871872, + "accuracy": 0.9995736529526766, + "total_bits": 44443904, "q_proj": { "group_size": { "8": 128 @@ -8946,8 +8946,8 @@ ], "model.layers.4.mlp": [ { - "accuracy": 0.944981537759304, - "total_bits": 29656128, + "accuracy": 0.964530885219574, + "total_bits": 93392016, "gate_proj": { "group_size": { "3": 64, @@ -8998,8 +8998,8 @@ } }, { - "accuracy": 0.9473154880106449, - "total_bits": 30901312, + "accuracy": 0.9656468406319618, + "total_bits": 96259216, "gate_proj": { "group_size": { "3": 64, @@ -9050,8 +9050,8 @@ } }, { - "accuracy": 0.9544729143381119, - "total_bits": 34456000, + "accuracy": 0.9698783811181784, + "total_bits": 107863424, "gate_proj": { "group_size": { "3": 64, @@ -9099,8 +9099,8 @@ } }, { - "accuracy": 0.9566175788640976, - "total_bits": 38584768, + "accuracy": 0.9712258372455835, + "total_bits": 120937856, "gate_proj": { "group_size": { "3": 64, @@ -9148,8 +9148,8 @@ } }, { - "accuracy": 0.9714861325919628, - "total_bits": 43342560, + "accuracy": 0.9819794353097677, + "total_bits": 136162224, "gate_proj": { "group_size": { "4": 128, @@ -9200,8 +9200,8 @@ } }, { - "accuracy": 0.9745184183120728, - "total_bits": 44242688, + "accuracy": 0.9835284482687712, + "total_bits": 139118208, "gate_proj": { "group_size": { "4": 32, @@ -9252,8 +9252,8 @@ } }, { - "accuracy": 0.9772712253034115, - "total_bits": 47520448, + "accuracy": 0.9851658642292023, + "total_bits": 149584176, "gate_proj": { "group_size": { "4": 32, @@ -9301,8 +9301,8 @@ } }, { - "accuracy": 0.9849097514525056, - "total_bits": 54270672, + "accuracy": 0.9903447218239307, + "total_bits": 170943088, "gate_proj": { "group_size": { "4": 128 @@ -9344,8 +9344,8 @@ } }, { - "accuracy": 0.9865038841962814, - "total_bits": 55072512, + "accuracy": 0.9912647157907486, + "total_bits": 173467264, "gate_proj": { "group_size": { "4": 32 @@ -9387,8 +9387,8 @@ } }, { - "accuracy": 0.9853217834606767, - "total_bits": 56187616, + "accuracy": 0.9907460622489452, + "total_bits": 176761776, "gate_proj": { "group_size": { "5": 128, @@ -9439,8 +9439,8 @@ } }, { - "accuracy": 0.9874983169138432, - "total_bits": 57087744, + "accuracy": 0.991934958845377, + "total_bits": 179717760, "gate_proj": { "group_size": { "5": 32, @@ -9491,8 +9491,8 @@ } }, { - "accuracy": 0.9924559579230845, - "total_bits": 69032672, + "accuracy": 0.9952533547766507, + "total_bits": 217361328, "gate_proj": { "group_size": { "6": 128, @@ -9543,8 +9543,8 @@ } }, { - "accuracy": 0.9937229603528976, - "total_bits": 69932800, + "accuracy": 0.9959681276232004, + "total_bits": 220317312, "gate_proj": { "group_size": { "6": 32, @@ -9595,8 +9595,8 @@ } }, { - "accuracy": 0.995773364789784, - "total_bits": 79551200, + "accuracy": 0.9973393860273063, + "total_bits": 250850224, "gate_proj": { "group_size": { "6": 128 @@ -9638,8 +9638,8 @@ } }, { - "accuracy": 0.9961159189697355, - "total_bits": 82937280, + "accuracy": 0.9976061794441193, + "total_bits": 262906272, "gate_proj": { "group_size": { "8": 128, @@ -9687,8 +9687,8 @@ } }, { - "accuracy": 0.9967866011429578, - "total_bits": 90277312, + "accuracy": 0.997991829412058, + "total_bits": 286106016, "gate_proj": { "group_size": { "8": 128, @@ -9733,8 +9733,8 @@ } }, { - "accuracy": 0.9988744062138721, - "total_bits": 105219520, + "accuracy": 0.9992650635540485, + "total_bits": 331981216, "gate_proj": { "group_size": { "8": 128 @@ -9775,8 +9775,8 @@ ], "model.layers.5.self_attn": [ { - "accuracy": 0.9634487144649029, - "total_bits": 4051584, + "accuracy": 0.9797085449099541, + "total_bits": 12046848, "q_proj": { "group_size": { "3": 64, @@ -9839,8 +9839,8 @@ } }, { - "accuracy": 0.9651153981685638, - "total_bits": 4199040, + "accuracy": 0.981008343398571, + "total_bits": 12325376, "q_proj": { "group_size": { "3": 64, @@ -9903,8 +9903,8 @@ } }, { - "accuracy": 0.9691105224192142, - "total_bits": 4293424, + "accuracy": 0.9833647850900888, + "total_bits": 12672960, "q_proj": { "group_size": { "3": 64, @@ -9967,8 +9967,8 @@ } }, { - "accuracy": 0.9727426841855049, - "total_bits": 5070816, + "accuracy": 0.9876857381314039, + "total_bits": 15056256, "q_proj": { "group_size": { "3": 64, @@ -10031,8 +10031,8 @@ } }, { - "accuracy": 0.9806992765516043, - "total_bits": 5958976, + "accuracy": 0.9899230813607574, + "total_bits": 17836288, "q_proj": { "group_size": { "4": 128, @@ -10095,8 +10095,8 @@ } }, { - "accuracy": 0.9805704932659864, - "total_bits": 5962896, + "accuracy": 0.9899922339245677, + "total_bits": 17832768, "q_proj": { "group_size": { "4": 128, @@ -10159,8 +10159,8 @@ } }, { - "accuracy": 0.9848807593807578, - "total_bits": 7531840, + "accuracy": 0.993687491863966, + "total_bits": 22423808, "q_proj": { "group_size": { "4": 128 @@ -10211,8 +10211,8 @@ } }, { - "accuracy": 0.9842194598168135, - "total_bits": 7535760, + "accuracy": 0.9938671756535769, + "total_bits": 22436672, "q_proj": { "group_size": { "4": 128 @@ -10263,8 +10263,8 @@ } }, { - "accuracy": 0.9876939542591572, - "total_bits": 7598368, + "accuracy": 0.9943577698431909, + "total_bits": 22623872, "q_proj": { "group_size": { "4": 64 @@ -10315,8 +10315,8 @@ } }, { - "accuracy": 0.9883776344358921, - "total_bits": 7707904, + "accuracy": 0.9946510489098728, + "total_bits": 22946816, "q_proj": { "group_size": { "4": 32 @@ -10367,8 +10367,8 @@ } }, { - "accuracy": 0.9911426873877645, - "total_bits": 7797904, + "accuracy": 0.994818109087646, + "total_bits": 23337792, "q_proj": { "group_size": { "5": 128, @@ -10431,8 +10431,8 @@ } }, { - "accuracy": 0.9917176300659776, - "total_bits": 7856416, + "accuracy": 0.9951027478091419, + "total_bits": 23303808, "q_proj": { "group_size": { "5": 64, @@ -10495,8 +10495,8 @@ } }, { - "accuracy": 0.9922088482417166, - "total_bits": 7950976, + "accuracy": 0.9956685532815754, + "total_bits": 23630336, "q_proj": { "group_size": { "5": 64, @@ -10556,8 +10556,8 @@ } }, { - "accuracy": 0.9928428521379828, - "total_bits": 8006912, + "accuracy": 0.9959396691992879, + "total_bits": 23872512, "q_proj": { "group_size": { "5": 32, @@ -10617,8 +10617,8 @@ } }, { - "accuracy": 0.9955793088302016, - "total_bits": 9727296, + "accuracy": 0.9976986155379564, + "total_bits": 29174016, "q_proj": { "group_size": { "6": 128, @@ -10678,8 +10678,8 @@ } }, { - "accuracy": 0.996340008918196, - "total_bits": 9841920, + "accuracy": 0.9979777017142624, + "total_bits": 29377536, "q_proj": { "group_size": { "6": 32, @@ -10739,8 +10739,8 @@ } }, { - "accuracy": 0.9964117147028446, - "total_bits": 11201856, + "accuracy": 0.9983391285641119, + "total_bits": 33433856, "q_proj": { "group_size": { "6": 128 @@ -10791,8 +10791,8 @@ } }, { - "accuracy": 0.9978703560773283, - "total_bits": 11607296, + "accuracy": 0.9988367387559265, + "total_bits": 34743296, "q_proj": { "group_size": { "6": 32 @@ -10843,8 +10843,8 @@ } }, { - "accuracy": 0.9990786031703465, - "total_bits": 14871872, + "accuracy": 0.9995475913165137, + "total_bits": 44443904, "q_proj": { "group_size": { "8": 128 @@ -10897,8 +10897,8 @@ ], "model.layers.5.mlp": [ { - "accuracy": 0.9473157450556755, - "total_bits": 29656128, + "accuracy": 0.9665445387363434, + "total_bits": 93392016, "gate_proj": { "group_size": { "3": 64, @@ -10949,8 +10949,8 @@ } }, { - "accuracy": 0.9495415687561035, - "total_bits": 30901312, + "accuracy": 0.9675517901778221, + "total_bits": 96259216, "gate_proj": { "group_size": { "3": 64, @@ -11001,8 +11001,8 @@ } }, { - "accuracy": 0.9551933743059635, - "total_bits": 34456000, + "accuracy": 0.9704751074314117, + "total_bits": 107863424, "gate_proj": { "group_size": { "3": 64, @@ -11050,8 +11050,8 @@ } }, { - "accuracy": 0.9569542519748211, - "total_bits": 38584768, + "accuracy": 0.9715234059840441, + "total_bits": 120937856, "gate_proj": { "group_size": { "3": 64, @@ -11099,8 +11099,8 @@ } }, { - "accuracy": 0.9728886485099792, - "total_bits": 43342560, + "accuracy": 0.9828967824578285, + "total_bits": 136162224, "gate_proj": { "group_size": { "4": 128, @@ -11151,8 +11151,8 @@ } }, { - "accuracy": 0.9756576213985682, - "total_bits": 44242688, + "accuracy": 0.9844599310308695, + "total_bits": 139118208, "gate_proj": { "group_size": { "4": 32, @@ -11203,8 +11203,8 @@ } }, { - "accuracy": 0.9778086096048355, - "total_bits": 47520448, + "accuracy": 0.9855620320886374, + "total_bits": 149584176, "gate_proj": { "group_size": { "4": 32, @@ -11252,8 +11252,8 @@ } }, { - "accuracy": 0.9858620781451464, - "total_bits": 54270672, + "accuracy": 0.9908638307824731, + "total_bits": 170943088, "gate_proj": { "group_size": { "4": 128 @@ -11295,8 +11295,8 @@ } }, { - "accuracy": 0.9871163554489613, - "total_bits": 55072512, + "accuracy": 0.9916770542040467, + "total_bits": 173467264, "gate_proj": { "group_size": { "4": 32 @@ -11338,8 +11338,8 @@ } }, { - "accuracy": 0.9861770961433649, - "total_bits": 56187616, + "accuracy": 0.9912218917161226, + "total_bits": 176761776, "gate_proj": { "group_size": { "5": 128, @@ -11390,8 +11390,8 @@ } }, { - "accuracy": 0.9881225265562534, - "total_bits": 57087744, + "accuracy": 0.9924233979545534, + "total_bits": 179717760, "gate_proj": { "group_size": { "5": 32, @@ -11442,8 +11442,8 @@ } }, { - "accuracy": 0.992890099529177, - "total_bits": 69032672, + "accuracy": 0.9954988393001258, + "total_bits": 217361328, "gate_proj": { "group_size": { "6": 128, @@ -11494,8 +11494,8 @@ } }, { - "accuracy": 0.9940621736459434, - "total_bits": 69932800, + "accuracy": 0.9961935267783701, + "total_bits": 220317312, "gate_proj": { "group_size": { "6": 32, @@ -11546,8 +11546,8 @@ } }, { - "accuracy": 0.9959961227141321, - "total_bits": 79551200, + "accuracy": 0.9974199496209621, + "total_bits": 250850224, "gate_proj": { "group_size": { "6": 128 @@ -11589,8 +11589,8 @@ } }, { - "accuracy": 0.9963513859547675, - "total_bits": 82937280, + "accuracy": 0.9977314700372517, + "total_bits": 262906272, "gate_proj": { "group_size": { "8": 128, @@ -11638,8 +11638,8 @@ } }, { - "accuracy": 0.996820350876078, - "total_bits": 90277312, + "accuracy": 0.9979731787461787, + "total_bits": 286106016, "gate_proj": { "group_size": { "8": 128, @@ -11684,8 +11684,8 @@ } }, { - "accuracy": 0.9989030649885535, - "total_bits": 105219520, + "accuracy": 0.9992820413899608, + "total_bits": 331981216, "gate_proj": { "group_size": { "8": 128 @@ -11726,8 +11726,8 @@ ], "model.layers.6.self_attn": [ { - "accuracy": 0.9677871391177177, - "total_bits": 4051584, + "accuracy": 0.9822309650480747, + "total_bits": 12046848, "q_proj": { "group_size": { "3": 64, @@ -11790,8 +11790,8 @@ } }, { - "accuracy": 0.9689766298979521, - "total_bits": 4199040, + "accuracy": 0.9831566084176302, + "total_bits": 12325376, "q_proj": { "group_size": { "3": 64, @@ -11854,8 +11854,8 @@ } }, { - "accuracy": 0.9725274182856083, - "total_bits": 4293424, + "accuracy": 0.9848638949915767, + "total_bits": 12672960, "q_proj": { "group_size": { "3": 64, @@ -11918,8 +11918,8 @@ } }, { - "accuracy": 0.9762999042868614, - "total_bits": 5070816, + "accuracy": 0.9880935745313764, + "total_bits": 15056256, "q_proj": { "group_size": { "3": 64, @@ -11982,8 +11982,8 @@ } }, { - "accuracy": 0.9801953807473183, - "total_bits": 5958976, + "accuracy": 0.9910181406885386, + "total_bits": 17836288, "q_proj": { "group_size": { "4": 128, @@ -12046,8 +12046,8 @@ } }, { - "accuracy": 0.9810614939779043, - "total_bits": 5962896, + "accuracy": 0.9911378072574735, + "total_bits": 17832768, "q_proj": { "group_size": { "4": 128, @@ -12110,8 +12110,8 @@ } }, { - "accuracy": 0.9830410778522491, - "total_bits": 7531840, + "accuracy": 0.9938726266846061, + "total_bits": 22423808, "q_proj": { "group_size": { "4": 128 @@ -12162,8 +12162,8 @@ } }, { - "accuracy": 0.9841166250407696, - "total_bits": 7535760, + "accuracy": 0.994075573515147, + "total_bits": 22436672, "q_proj": { "group_size": { "4": 128 @@ -12214,8 +12214,8 @@ } }, { - "accuracy": 0.9870422091335058, - "total_bits": 7598368, + "accuracy": 0.994407317135483, + "total_bits": 22623872, "q_proj": { "group_size": { "4": 64 @@ -12266,8 +12266,8 @@ } }, { - "accuracy": 0.9869954623281956, - "total_bits": 7707904, + "accuracy": 0.9948621308431029, + "total_bits": 22946816, "q_proj": { "group_size": { "4": 32 @@ -12318,8 +12318,8 @@ } }, { - "accuracy": 0.9907674258574843, - "total_bits": 7797904, + "accuracy": 0.9955409793183208, + "total_bits": 23337792, "q_proj": { "group_size": { "5": 128, @@ -12382,8 +12382,8 @@ } }, { - "accuracy": 0.9889425421133637, - "total_bits": 7856416, + "accuracy": 0.9956872593611479, + "total_bits": 23303808, "q_proj": { "group_size": { "5": 64, @@ -12446,8 +12446,8 @@ } }, { - "accuracy": 0.9883685754612088, - "total_bits": 7950976, + "accuracy": 0.9959687143564224, + "total_bits": 23630336, "q_proj": { "group_size": { "5": 64, @@ -12507,8 +12507,8 @@ } }, { - "accuracy": 0.9906774684786797, - "total_bits": 8006912, + "accuracy": 0.9964766607154161, + "total_bits": 23872512, "q_proj": { "group_size": { "5": 32, @@ -12568,8 +12568,8 @@ } }, { - "accuracy": 0.9957364844158292, - "total_bits": 9727296, + "accuracy": 0.9979342515580356, + "total_bits": 29174016, "q_proj": { "group_size": { "6": 128, @@ -12629,8 +12629,8 @@ } }, { - "accuracy": 0.9961910455022007, - "total_bits": 9841920, + "accuracy": 0.9982180697843432, + "total_bits": 29377536, "q_proj": { "group_size": { "6": 32, @@ -12690,8 +12690,8 @@ } }, { - "accuracy": 0.9964089696295559, - "total_bits": 11201856, + "accuracy": 0.9984661155613139, + "total_bits": 33433856, "q_proj": { "group_size": { "6": 128 @@ -12742,8 +12742,8 @@ } }, { - "accuracy": 0.9975835788063705, - "total_bits": 11607296, + "accuracy": 0.9988858799915761, + "total_bits": 34743296, "q_proj": { "group_size": { "6": 32 @@ -12794,8 +12794,8 @@ } }, { - "accuracy": 0.9990214385325089, - "total_bits": 14871872, + "accuracy": 0.9995953500911128, + "total_bits": 44443904, "q_proj": { "group_size": { "8": 128 @@ -12848,8 +12848,8 @@ ], "model.layers.6.mlp": [ { - "accuracy": 0.9464025646448135, - "total_bits": 29656128, + "accuracy": 0.9658905267715454, + "total_bits": 93392016, "gate_proj": { "group_size": { "3": 64, @@ -12900,8 +12900,8 @@ } }, { - "accuracy": 0.9487528428435326, - "total_bits": 30901312, + "accuracy": 0.9669794477522373, + "total_bits": 96259216, "gate_proj": { "group_size": { "3": 64, @@ -12952,8 +12952,8 @@ } }, { - "accuracy": 0.9553916938602924, - "total_bits": 34456000, + "accuracy": 0.9712447710335255, + "total_bits": 107863424, "gate_proj": { "group_size": { "3": 64, @@ -13001,8 +13001,8 @@ } }, { - "accuracy": 0.9574474431574345, - "total_bits": 38584768, + "accuracy": 0.9724618531763554, + "total_bits": 120937856, "gate_proj": { "group_size": { "3": 64, @@ -13050,8 +13050,8 @@ } }, { - "accuracy": 0.972553975880146, - "total_bits": 43342560, + "accuracy": 0.9828744474798441, + "total_bits": 136162224, "gate_proj": { "group_size": { "4": 128, @@ -13102,8 +13102,8 @@ } }, { - "accuracy": 0.9753192290663719, - "total_bits": 44242688, + "accuracy": 0.9843734856694937, + "total_bits": 139118208, "gate_proj": { "group_size": { "4": 32, @@ -13154,8 +13154,8 @@ } }, { - "accuracy": 0.9779228679835796, - "total_bits": 47520448, + "accuracy": 0.985878842882812, + "total_bits": 149584176, "gate_proj": { "group_size": { "4": 32, @@ -13203,8 +13203,8 @@ } }, { - "accuracy": 0.9854622259736061, - "total_bits": 54270672, + "accuracy": 0.9910057932138443, + "total_bits": 170943088, "gate_proj": { "group_size": { "4": 128 @@ -13246,8 +13246,8 @@ } }, { - "accuracy": 0.9869041200727224, - "total_bits": 55072512, + "accuracy": 0.9917910546064377, + "total_bits": 173467264, "gate_proj": { "group_size": { "4": 32 @@ -13289,8 +13289,8 @@ } }, { - "accuracy": 0.9858766980469227, - "total_bits": 56187616, + "accuracy": 0.991222495213151, + "total_bits": 176761776, "gate_proj": { "group_size": { "5": 128, @@ -13341,8 +13341,8 @@ } }, { - "accuracy": 0.9879383649677038, - "total_bits": 57087744, + "accuracy": 0.992395996581763, + "total_bits": 179717760, "gate_proj": { "group_size": { "5": 32, @@ -13393,8 +13393,8 @@ } }, { - "accuracy": 0.9927426120266318, - "total_bits": 69032672, + "accuracy": 0.995514387730509, + "total_bits": 217361328, "gate_proj": { "group_size": { "6": 128, @@ -13445,8 +13445,8 @@ } }, { - "accuracy": 0.9939632629975677, - "total_bits": 69932800, + "accuracy": 0.9961990755982697, + "total_bits": 220317312, "gate_proj": { "group_size": { "6": 32, @@ -13497,8 +13497,8 @@ } }, { - "accuracy": 0.995909676887095, - "total_bits": 79551200, + "accuracy": 0.9975077514536679, + "total_bits": 250850224, "gate_proj": { "group_size": { "6": 128 @@ -13540,8 +13540,8 @@ } }, { - "accuracy": 0.996296726167202, - "total_bits": 82937280, + "accuracy": 0.9977307666558772, + "total_bits": 262906272, "gate_proj": { "group_size": { "8": 128, @@ -13589,8 +13589,8 @@ } }, { - "accuracy": 0.9968657279387116, - "total_bits": 90277312, + "accuracy": 0.9980514174094424, + "total_bits": 286106016, "gate_proj": { "group_size": { "8": 128, @@ -13635,8 +13635,8 @@ } }, { - "accuracy": 0.9989077572245151, - "total_bits": 105219520, + "accuracy": 0.9993329935823567, + "total_bits": 331981216, "gate_proj": { "group_size": { "8": 128 @@ -13677,8 +13677,8 @@ ], "model.layers.7.self_attn": [ { - "accuracy": 0.9291665181517601, - "total_bits": 4051584, + "accuracy": 0.9817934613674879, + "total_bits": 12046848, "q_proj": { "group_size": { "3": 64, @@ -13741,8 +13741,8 @@ } }, { - "accuracy": 0.95192801207304, - "total_bits": 4199040, + "accuracy": 0.9836168773472309, + "total_bits": 12325376, "q_proj": { "group_size": { "3": 64, @@ -13805,8 +13805,8 @@ } }, { - "accuracy": 0.9555695615708828, - "total_bits": 4293424, + "accuracy": 0.9857309749349952, + "total_bits": 12672960, "q_proj": { "group_size": { "3": 64, @@ -13869,8 +13869,8 @@ } }, { - "accuracy": 0.9606849290430546, - "total_bits": 5070816, + "accuracy": 0.9898735983297229, + "total_bits": 15056256, "q_proj": { "group_size": { "3": 64, @@ -13933,8 +13933,8 @@ } }, { - "accuracy": 0.9785683583468199, - "total_bits": 5958976, + "accuracy": 0.9912841729819775, + "total_bits": 17836288, "q_proj": { "group_size": { "4": 128, @@ -13997,8 +13997,8 @@ } }, { - "accuracy": 0.9786245785653591, - "total_bits": 5962896, + "accuracy": 0.9914414696395397, + "total_bits": 17832768, "q_proj": { "group_size": { "4": 128, @@ -14061,8 +14061,8 @@ } }, { - "accuracy": 0.9839308448135853, - "total_bits": 7531840, + "accuracy": 0.9945913958363235, + "total_bits": 22423808, "q_proj": { "group_size": { "4": 128 @@ -14113,8 +14113,8 @@ } }, { - "accuracy": 0.9838340990245342, - "total_bits": 7535760, + "accuracy": 0.9948036088608205, + "total_bits": 22436672, "q_proj": { "group_size": { "4": 128 @@ -14165,8 +14165,8 @@ } }, { - "accuracy": 0.9835375174880028, - "total_bits": 7598368, + "accuracy": 0.9951845020987093, + "total_bits": 22623872, "q_proj": { "group_size": { "4": 64 @@ -14217,8 +14217,8 @@ } }, { - "accuracy": 0.9863764280453324, - "total_bits": 7707904, + "accuracy": 0.9955280623398721, + "total_bits": 22946816, "q_proj": { "group_size": { "4": 32 @@ -14269,8 +14269,8 @@ } }, { - "accuracy": 0.9885445488616824, - "total_bits": 7797904, + "accuracy": 0.9956175033003092, + "total_bits": 23337792, "q_proj": { "group_size": { "5": 128, @@ -14333,8 +14333,8 @@ } }, { - "accuracy": 0.9901036657392979, - "total_bits": 7856416, + "accuracy": 0.9958917181938887, + "total_bits": 23303808, "q_proj": { "group_size": { "5": 64, @@ -14397,8 +14397,8 @@ } }, { - "accuracy": 0.9901812877506018, - "total_bits": 7950976, + "accuracy": 0.996321314945817, + "total_bits": 23630336, "q_proj": { "group_size": { "5": 64, @@ -14458,8 +14458,8 @@ } }, { - "accuracy": 0.9917040690779686, - "total_bits": 8006912, + "accuracy": 0.9966362807899714, + "total_bits": 23872512, "q_proj": { "group_size": { "5": 32, @@ -14519,8 +14519,8 @@ } }, { - "accuracy": 0.9947202098555863, - "total_bits": 9727296, + "accuracy": 0.9980035650078207, + "total_bits": 29174016, "q_proj": { "group_size": { "6": 128, @@ -14580,8 +14580,8 @@ } }, { - "accuracy": 0.9954890562221408, - "total_bits": 9841920, + "accuracy": 0.9983131164917722, + "total_bits": 29377536, "q_proj": { "group_size": { "6": 32, @@ -14641,8 +14641,8 @@ } }, { - "accuracy": 0.9957401473075151, - "total_bits": 11201856, + "accuracy": 0.9986020227661356, + "total_bits": 33433856, "q_proj": { "group_size": { "6": 128 @@ -14693,8 +14693,8 @@ } }, { - "accuracy": 0.9971228996291757, - "total_bits": 11607296, + "accuracy": 0.9991119984188117, + "total_bits": 34743296, "q_proj": { "group_size": { "6": 32 @@ -14745,8 +14745,8 @@ } }, { - "accuracy": 0.9989357761805877, - "total_bits": 14871872, + "accuracy": 0.9996108976774849, + "total_bits": 44443904, "q_proj": { "group_size": { "8": 128 @@ -14799,8 +14799,8 @@ ], "model.layers.7.mlp": [ { - "accuracy": 0.9435533210635185, - "total_bits": 29656128, + "accuracy": 0.9661630131304264, + "total_bits": 93392016, "gate_proj": { "group_size": { "3": 64, @@ -14851,8 +14851,8 @@ } }, { - "accuracy": 0.9459276646375656, - "total_bits": 30901312, + "accuracy": 0.9671835377812386, + "total_bits": 96259216, "gate_proj": { "group_size": { "3": 64, @@ -14903,8 +14903,8 @@ } }, { - "accuracy": 0.9534953087568283, - "total_bits": 34456000, + "accuracy": 0.9717093147337437, + "total_bits": 107863424, "gate_proj": { "group_size": { "3": 64, @@ -14952,8 +14952,8 @@ } }, { - "accuracy": 0.9558756239712238, - "total_bits": 38584768, + "accuracy": 0.9730532318353653, + "total_bits": 120937856, "gate_proj": { "group_size": { "3": 64, @@ -15001,8 +15001,8 @@ } }, { - "accuracy": 0.9703601785004139, - "total_bits": 43342560, + "accuracy": 0.9826632346957922, + "total_bits": 136162224, "gate_proj": { "group_size": { "4": 128, @@ -15053,8 +15053,8 @@ } }, { - "accuracy": 0.9736221209168434, - "total_bits": 44242688, + "accuracy": 0.9843854084610939, + "total_bits": 139118208, "gate_proj": { "group_size": { "4": 32, @@ -15105,8 +15105,8 @@ } }, { - "accuracy": 0.9764794055372477, - "total_bits": 47520448, + "accuracy": 0.9860421344637871, + "total_bits": 149584176, "gate_proj": { "group_size": { "4": 32, @@ -15154,8 +15154,8 @@ } }, { - "accuracy": 0.9843061491847038, - "total_bits": 54270672, + "accuracy": 0.9908044775947928, + "total_bits": 170943088, "gate_proj": { "group_size": { "4": 128 @@ -15197,8 +15197,8 @@ } }, { - "accuracy": 0.9859360186383128, - "total_bits": 55072512, + "accuracy": 0.9917597165331244, + "total_bits": 173467264, "gate_proj": { "group_size": { "4": 32 @@ -15240,8 +15240,8 @@ } }, { - "accuracy": 0.9846695018932223, - "total_bits": 56187616, + "accuracy": 0.9910596255213022, + "total_bits": 176761776, "gate_proj": { "group_size": { "5": 128, @@ -15292,8 +15292,8 @@ } }, { - "accuracy": 0.9870050325989723, - "total_bits": 57087744, + "accuracy": 0.9923602067865431, + "total_bits": 179717760, "gate_proj": { "group_size": { "5": 32, @@ -15344,8 +15344,8 @@ } }, { - "accuracy": 0.9921077564358711, - "total_bits": 69032672, + "accuracy": 0.9954051352106035, + "total_bits": 217361328, "gate_proj": { "group_size": { "6": 128, @@ -15396,8 +15396,8 @@ } }, { - "accuracy": 0.993465397041291, - "total_bits": 69932800, + "accuracy": 0.9961705161258578, + "total_bits": 220317312, "gate_proj": { "group_size": { "6": 32, @@ -15448,8 +15448,8 @@ } }, { - "accuracy": 0.9955712254159153, - "total_bits": 79551200, + "accuracy": 0.9974461845122278, + "total_bits": 250850224, "gate_proj": { "group_size": { "6": 128 @@ -15491,8 +15491,8 @@ } }, { - "accuracy": 0.9959376780316234, - "total_bits": 82937280, + "accuracy": 0.9976750721689314, + "total_bits": 262906272, "gate_proj": { "group_size": { "8": 128, @@ -15540,8 +15540,8 @@ } }, { - "accuracy": 0.9966191961430013, - "total_bits": 90277312, + "accuracy": 0.998048037989065, + "total_bits": 286106016, "gate_proj": { "group_size": { "8": 128, @@ -15586,8 +15586,8 @@ } }, { - "accuracy": 0.9988139675697312, - "total_bits": 105219520, + "accuracy": 0.9993136781267822, + "total_bits": 331981216, "gate_proj": { "group_size": { "8": 128 @@ -15628,8 +15628,8 @@ ], "model.layers.8.self_attn": [ { - "accuracy": 0.9535798579454422, - "total_bits": 4051584, + "accuracy": 0.9785557892173529, + "total_bits": 12046848, "q_proj": { "group_size": { "3": 64, @@ -15692,8 +15692,8 @@ } }, { - "accuracy": 0.9567788764834404, - "total_bits": 4199040, + "accuracy": 0.9792898017913103, + "total_bits": 12325376, "q_proj": { "group_size": { "3": 64, @@ -15756,8 +15756,8 @@ } }, { - "accuracy": 0.9603584669530392, - "total_bits": 4293424, + "accuracy": 0.9807907361537218, + "total_bits": 12672960, "q_proj": { "group_size": { "3": 64, @@ -15820,8 +15820,8 @@ } }, { - "accuracy": 0.9675993286073208, - "total_bits": 5070816, + "accuracy": 0.9851897191256285, + "total_bits": 15056256, "q_proj": { "group_size": { "3": 64, @@ -15884,8 +15884,8 @@ } }, { - "accuracy": 0.9749189633876085, - "total_bits": 5958976, + "accuracy": 0.9891376812011003, + "total_bits": 17836288, "q_proj": { "group_size": { "4": 128, @@ -15948,8 +15948,8 @@ } }, { - "accuracy": 0.9753494802862406, - "total_bits": 5962896, + "accuracy": 0.9892689185217023, + "total_bits": 17832768, "q_proj": { "group_size": { "4": 128, @@ -16012,8 +16012,8 @@ } }, { - "accuracy": 0.9801577106118202, - "total_bits": 7531840, + "accuracy": 0.992960664909333, + "total_bits": 22423808, "q_proj": { "group_size": { "4": 128 @@ -16064,8 +16064,8 @@ } }, { - "accuracy": 0.9805457498878241, - "total_bits": 7535760, + "accuracy": 0.9931711829267442, + "total_bits": 22436672, "q_proj": { "group_size": { "4": 128 @@ -16116,8 +16116,8 @@ } }, { - "accuracy": 0.9847986474633217, - "total_bits": 7598368, + "accuracy": 0.9938573935069144, + "total_bits": 22623872, "q_proj": { "group_size": { "4": 64 @@ -16168,8 +16168,8 @@ } }, { - "accuracy": 0.9857409968972206, - "total_bits": 7707904, + "accuracy": 0.9942574431188405, + "total_bits": 22946816, "q_proj": { "group_size": { "4": 32 @@ -16220,8 +16220,8 @@ } }, { - "accuracy": 0.9854198498651385, - "total_bits": 7797904, + "accuracy": 0.9945622179657221, + "total_bits": 23337792, "q_proj": { "group_size": { "5": 128, @@ -16284,8 +16284,8 @@ } }, { - "accuracy": 0.9864915767684579, - "total_bits": 7856416, + "accuracy": 0.9947910481132567, + "total_bits": 23303808, "q_proj": { "group_size": { "5": 64, @@ -16348,8 +16348,8 @@ } }, { - "accuracy": 0.986890478990972, - "total_bits": 7950976, + "accuracy": 0.9951277710497379, + "total_bits": 23630336, "q_proj": { "group_size": { "5": 64, @@ -16409,8 +16409,8 @@ } }, { - "accuracy": 0.9878804571926594, - "total_bits": 8006912, + "accuracy": 0.9952990976162255, + "total_bits": 23872512, "q_proj": { "group_size": { "5": 32, @@ -16470,8 +16470,8 @@ } }, { - "accuracy": 0.9920131107792258, - "total_bits": 9727296, + "accuracy": 0.9973937598988414, + "total_bits": 29174016, "q_proj": { "group_size": { "6": 128, @@ -16531,8 +16531,8 @@ } }, { - "accuracy": 0.994783956091851, - "total_bits": 9841920, + "accuracy": 0.9977938737720251, + "total_bits": 29377536, "q_proj": { "group_size": { "6": 32, @@ -16592,8 +16592,8 @@ } }, { - "accuracy": 0.9926207619719207, - "total_bits": 11201856, + "accuracy": 0.9982025797944516, + "total_bits": 33433856, "q_proj": { "group_size": { "6": 128 @@ -16644,8 +16644,8 @@ } }, { - "accuracy": 0.9964182772673666, - "total_bits": 11607296, + "accuracy": 0.9986881435615942, + "total_bits": 34743296, "q_proj": { "group_size": { "6": 32 @@ -16696,8 +16696,8 @@ } }, { - "accuracy": 0.9987832672195509, - "total_bits": 14871872, + "accuracy": 0.9995204001606908, + "total_bits": 44443904, "q_proj": { "group_size": { "8": 128 @@ -16750,8 +16750,8 @@ ], "model.layers.8.mlp": [ { - "accuracy": 0.9523984827101231, - "total_bits": 29656128, + "accuracy": 0.9675015024840832, + "total_bits": 93392016, "gate_proj": { "group_size": { "3": 64, @@ -16802,8 +16802,8 @@ } }, { - "accuracy": 0.9545088484883308, - "total_bits": 30901312, + "accuracy": 0.9684530273079872, + "total_bits": 96259216, "gate_proj": { "group_size": { "3": 64, @@ -16854,8 +16854,8 @@ } }, { - "accuracy": 0.9609974473714828, - "total_bits": 34456000, + "accuracy": 0.9727552440017462, + "total_bits": 107863424, "gate_proj": { "group_size": { "3": 64, @@ -16903,8 +16903,8 @@ } }, { - "accuracy": 0.9629328325390816, - "total_bits": 38584768, + "accuracy": 0.9740058407187462, + "total_bits": 120937856, "gate_proj": { "group_size": { "3": 64, @@ -16952,8 +16952,8 @@ } }, { - "accuracy": 0.9757286049425602, - "total_bits": 43342560, + "accuracy": 0.9835954252630472, + "total_bits": 136162224, "gate_proj": { "group_size": { "4": 128, @@ -17004,8 +17004,8 @@ } }, { - "accuracy": 0.9779635332524776, - "total_bits": 44242688, + "accuracy": 0.9850423512980342, + "total_bits": 139118208, "gate_proj": { "group_size": { "4": 32, @@ -17056,8 +17056,8 @@ } }, { - "accuracy": 0.9805152211338282, - "total_bits": 47520448, + "accuracy": 0.9866323973983526, + "total_bits": 149584176, "gate_proj": { "group_size": { "4": 32, @@ -17105,8 +17105,8 @@ } }, { - "accuracy": 0.9869852541014552, - "total_bits": 54270672, + "accuracy": 0.9914277009665966, + "total_bits": 170943088, "gate_proj": { "group_size": { "4": 128 @@ -17148,8 +17148,8 @@ } }, { - "accuracy": 0.9881956651806831, - "total_bits": 55072512, + "accuracy": 0.992175011895597, + "total_bits": 173467264, "gate_proj": { "group_size": { "4": 32 @@ -17191,8 +17191,8 @@ } }, { - "accuracy": 0.9875301346182823, - "total_bits": 56187616, + "accuracy": 0.9915884742513299, + "total_bits": 176761776, "gate_proj": { "group_size": { "5": 128, @@ -17243,8 +17243,8 @@ } }, { - "accuracy": 0.989203886128962, - "total_bits": 57087744, + "accuracy": 0.9927049563266337, + "total_bits": 179717760, "gate_proj": { "group_size": { "5": 32, @@ -17295,8 +17295,8 @@ } }, { - "accuracy": 0.9936269242316484, - "total_bits": 69032672, + "accuracy": 0.9956925283186138, + "total_bits": 217361328, "gate_proj": { "group_size": { "6": 128, @@ -17347,8 +17347,8 @@ } }, { - "accuracy": 0.9946013330481946, - "total_bits": 69932800, + "accuracy": 0.996348132379353, + "total_bits": 220317312, "gate_proj": { "group_size": { "6": 32, @@ -17399,8 +17399,8 @@ } }, { - "accuracy": 0.9963683553505689, - "total_bits": 79551200, + "accuracy": 0.9976248391903937, + "total_bits": 250850224, "gate_proj": { "group_size": { "6": 128 @@ -17442,8 +17442,8 @@ } }, { - "accuracy": 0.9967480967752635, - "total_bits": 82937280, + "accuracy": 0.9978192530106753, + "total_bits": 262906272, "gate_proj": { "group_size": { "8": 128, @@ -17491,8 +17491,8 @@ } }, { - "accuracy": 0.9973138570785522, - "total_bits": 90277312, + "accuracy": 0.9981697353068739, + "total_bits": 286106016, "gate_proj": { "group_size": { "8": 128, @@ -17537,8 +17537,8 @@ } }, { - "accuracy": 0.9990280251367949, - "total_bits": 105219520, + "accuracy": 0.9993592473329045, + "total_bits": 331981216, "gate_proj": { "group_size": { "8": 128 @@ -17579,8 +17579,8 @@ ], "model.layers.9.self_attn": [ { - "accuracy": 0.9431938789784908, - "total_bits": 4051584, + "accuracy": 0.980949530377984, + "total_bits": 12046848, "q_proj": { "group_size": { "3": 64, @@ -17643,8 +17643,8 @@ } }, { - "accuracy": 0.942477386444807, - "total_bits": 4199040, + "accuracy": 0.9816097822040319, + "total_bits": 12325376, "q_proj": { "group_size": { "3": 64, @@ -17707,8 +17707,8 @@ } }, { - "accuracy": 0.9446651823818684, - "total_bits": 4293424, + "accuracy": 0.983317194506526, + "total_bits": 12672960, "q_proj": { "group_size": { "3": 64, @@ -17771,8 +17771,8 @@ } }, { - "accuracy": 0.9475866369903088, - "total_bits": 5070816, + "accuracy": 0.9875798439607024, + "total_bits": 15056256, "q_proj": { "group_size": { "3": 64, @@ -17835,8 +17835,8 @@ } }, { - "accuracy": 0.9684545509517193, - "total_bits": 5958976, + "accuracy": 0.9905973458662629, + "total_bits": 17836288, "q_proj": { "group_size": { "4": 128, @@ -17899,8 +17899,8 @@ } }, { - "accuracy": 0.9692513011395931, - "total_bits": 5962896, + "accuracy": 0.9906548298895359, + "total_bits": 17832768, "q_proj": { "group_size": { "4": 128, @@ -17963,8 +17963,8 @@ } }, { - "accuracy": 0.9790745433419943, - "total_bits": 7531840, + "accuracy": 0.9944175197742879, + "total_bits": 22423808, "q_proj": { "group_size": { "4": 128 @@ -18015,8 +18015,8 @@ } }, { - "accuracy": 0.9796556793153286, - "total_bits": 7535760, + "accuracy": 0.9945195005275309, + "total_bits": 22436672, "q_proj": { "group_size": { "4": 128 @@ -18067,8 +18067,8 @@ } }, { - "accuracy": 0.9834481533616781, - "total_bits": 7598368, + "accuracy": 0.9947983045130968, + "total_bits": 22623872, "q_proj": { "group_size": { "4": 64 @@ -18119,8 +18119,8 @@ } }, { - "accuracy": 0.9855514811351895, - "total_bits": 7707904, + "accuracy": 0.9951676898635924, + "total_bits": 22946816, "q_proj": { "group_size": { "4": 32 @@ -18171,8 +18171,8 @@ } }, { - "accuracy": 0.9852889999747276, - "total_bits": 7797904, + "accuracy": 0.9952615392394364, + "total_bits": 23337792, "q_proj": { "group_size": { "5": 128, @@ -18235,8 +18235,8 @@ } }, { - "accuracy": 0.9871742110699415, - "total_bits": 7856416, + "accuracy": 0.9954822142608464, + "total_bits": 23303808, "q_proj": { "group_size": { "5": 64, @@ -18299,8 +18299,8 @@ } }, { - "accuracy": 0.9883428532630205, - "total_bits": 7950976, + "accuracy": 0.9958796505816281, + "total_bits": 23630336, "q_proj": { "group_size": { "5": 64, @@ -18360,8 +18360,8 @@ } }, { - "accuracy": 0.9905433086678386, - "total_bits": 8006912, + "accuracy": 0.9961906985845417, + "total_bits": 23872512, "q_proj": { "group_size": { "5": 32, @@ -18421,8 +18421,8 @@ } }, { - "accuracy": 0.9911938859149814, - "total_bits": 9727296, + "accuracy": 0.997784714680165, + "total_bits": 29174016, "q_proj": { "group_size": { "6": 128, @@ -18482,8 +18482,8 @@ } }, { - "accuracy": 0.9948418377898633, - "total_bits": 9841920, + "accuracy": 0.9980635978281498, + "total_bits": 29377536, "q_proj": { "group_size": { "6": 32, @@ -18543,8 +18543,8 @@ } }, { - "accuracy": 0.9943175842054188, - "total_bits": 11201856, + "accuracy": 0.9985398501157761, + "total_bits": 33433856, "q_proj": { "group_size": { "6": 128 @@ -18595,8 +18595,8 @@ } }, { - "accuracy": 0.9963090608362108, - "total_bits": 11607296, + "accuracy": 0.9989193498622626, + "total_bits": 34743296, "q_proj": { "group_size": { "6": 32 @@ -18647,8 +18647,8 @@ } }, { - "accuracy": 0.9984725409885868, - "total_bits": 14871872, + "accuracy": 0.9995929627912119, + "total_bits": 44443904, "q_proj": { "group_size": { "8": 128 @@ -18701,8 +18701,8 @@ ], "model.layers.9.mlp": [ { - "accuracy": 0.9518498852849007, - "total_bits": 29656128, + "accuracy": 0.9688989818096161, + "total_bits": 93392016, "gate_proj": { "group_size": { "3": 64, @@ -18753,8 +18753,8 @@ } }, { - "accuracy": 0.9543508887290955, - "total_bits": 30901312, + "accuracy": 0.9697781074792147, + "total_bits": 96259216, "gate_proj": { "group_size": { "3": 64, @@ -18805,8 +18805,8 @@ } }, { - "accuracy": 0.960621777921915, - "total_bits": 34456000, + "accuracy": 0.9741190243512392, + "total_bits": 107863424, "gate_proj": { "group_size": { "3": 64, @@ -18854,8 +18854,8 @@ } }, { - "accuracy": 0.9625387638807297, - "total_bits": 38584768, + "accuracy": 0.9753765091300011, + "total_bits": 120937856, "gate_proj": { "group_size": { "3": 64, @@ -18903,8 +18903,8 @@ } }, { - "accuracy": 0.975132767111063, - "total_bits": 43342560, + "accuracy": 0.9841837249696255, + "total_bits": 136162224, "gate_proj": { "group_size": { "4": 128, @@ -18955,8 +18955,8 @@ } }, { - "accuracy": 0.9776170738041401, - "total_bits": 44242688, + "accuracy": 0.9855790212750435, + "total_bits": 139118208, "gate_proj": { "group_size": { "4": 32, @@ -19007,8 +19007,8 @@ } }, { - "accuracy": 0.9800954759120941, - "total_bits": 47520448, + "accuracy": 0.9872328266501427, + "total_bits": 149584176, "gate_proj": { "group_size": { "4": 32, @@ -19056,8 +19056,8 @@ } }, { - "accuracy": 0.9867881629616022, - "total_bits": 54270672, + "accuracy": 0.9917415557429194, + "total_bits": 170943088, "gate_proj": { "group_size": { "4": 128 @@ -19099,8 +19099,8 @@ } }, { - "accuracy": 0.9880188480019569, - "total_bits": 55072512, + "accuracy": 0.9924838095903397, + "total_bits": 173467264, "gate_proj": { "group_size": { "4": 32 @@ -19142,8 +19142,8 @@ } }, { - "accuracy": 0.9870889727026224, - "total_bits": 56187616, + "accuracy": 0.9918960658833385, + "total_bits": 176761776, "gate_proj": { "group_size": { "5": 128, @@ -19194,8 +19194,8 @@ } }, { - "accuracy": 0.9889920046553016, - "total_bits": 57087744, + "accuracy": 0.9929614635184407, + "total_bits": 179717760, "gate_proj": { "group_size": { "5": 32, @@ -19246,8 +19246,8 @@ } }, { - "accuracy": 0.9933229382149875, - "total_bits": 69032672, + "accuracy": 0.9958424451760948, + "total_bits": 217361328, "gate_proj": { "group_size": { "6": 128, @@ -19298,8 +19298,8 @@ } }, { - "accuracy": 0.994465074967593, - "total_bits": 69932800, + "accuracy": 0.9964762812014669, + "total_bits": 220317312, "gate_proj": { "group_size": { "6": 32, @@ -19350,8 +19350,8 @@ } }, { - "accuracy": 0.9962036833167076, - "total_bits": 79551200, + "accuracy": 0.9977118130773306, + "total_bits": 250850224, "gate_proj": { "group_size": { "6": 128 @@ -19393,8 +19393,8 @@ } }, { - "accuracy": 0.9965238117147237, - "total_bits": 82937280, + "accuracy": 0.997892533428967, + "total_bits": 262906272, "gate_proj": { "group_size": { "8": 128, @@ -19442,8 +19442,8 @@ } }, { - "accuracy": 0.9970647210720927, - "total_bits": 90277312, + "accuracy": 0.9982592724263668, + "total_bits": 286106016, "gate_proj": { "group_size": { "8": 128, @@ -19488,8 +19488,8 @@ } }, { - "accuracy": 0.9989702712045982, - "total_bits": 105219520, + "accuracy": 0.9993866681470536, + "total_bits": 331981216, "gate_proj": { "group_size": { "8": 128 @@ -19530,8 +19530,8 @@ ], "model.layers.10.self_attn": [ { - "accuracy": 0.9538821652531624, - "total_bits": 4051584, + "accuracy": 0.9782167710363865, + "total_bits": 12046848, "q_proj": { "group_size": { "3": 64, @@ -19594,8 +19594,8 @@ } }, { - "accuracy": 0.9584533907473087, - "total_bits": 4199040, + "accuracy": 0.9792514238506556, + "total_bits": 12325376, "q_proj": { "group_size": { "3": 64, @@ -19658,8 +19658,8 @@ } }, { - "accuracy": 0.964867889881134, - "total_bits": 4293424, + "accuracy": 0.9815011247992516, + "total_bits": 12672960, "q_proj": { "group_size": { "3": 64, @@ -19722,8 +19722,8 @@ } }, { - "accuracy": 0.9716326706111431, - "total_bits": 5070816, + "accuracy": 0.9869792461395264, + "total_bits": 15056256, "q_proj": { "group_size": { "3": 64, @@ -19786,8 +19786,8 @@ } }, { - "accuracy": 0.9766751136630774, - "total_bits": 5958976, + "accuracy": 0.9891741955652833, + "total_bits": 17836288, "q_proj": { "group_size": { "4": 128, @@ -19850,8 +19850,8 @@ } }, { - "accuracy": 0.9771394394338131, - "total_bits": 5962896, + "accuracy": 0.9893710343167186, + "total_bits": 17832768, "q_proj": { "group_size": { "4": 128, @@ -19914,8 +19914,8 @@ } }, { - "accuracy": 0.981711508706212, - "total_bits": 7531840, + "accuracy": 0.9933416000567377, + "total_bits": 22423808, "q_proj": { "group_size": { "4": 128 @@ -19966,8 +19966,8 @@ } }, { - "accuracy": 0.9819993413984776, - "total_bits": 7535760, + "accuracy": 0.9935925384052098, + "total_bits": 22436672, "q_proj": { "group_size": { "4": 128 @@ -20018,8 +20018,8 @@ } }, { - "accuracy": 0.9853426208719611, - "total_bits": 7598368, + "accuracy": 0.9942276077345014, + "total_bits": 22623872, "q_proj": { "group_size": { "4": 64 @@ -20070,8 +20070,8 @@ } }, { - "accuracy": 0.9854864981025457, - "total_bits": 7707904, + "accuracy": 0.9945281655527651, + "total_bits": 22946816, "q_proj": { "group_size": { "4": 32 @@ -20122,8 +20122,8 @@ } }, { - "accuracy": 0.9890922848135233, - "total_bits": 7797904, + "accuracy": 0.9946372536942363, + "total_bits": 23337792, "q_proj": { "group_size": { "5": 128, @@ -20186,8 +20186,8 @@ } }, { - "accuracy": 0.98984903562814, - "total_bits": 7856416, + "accuracy": 0.9948680996894836, + "total_bits": 23303808, "q_proj": { "group_size": { "5": 64, @@ -20250,8 +20250,8 @@ } }, { - "accuracy": 0.990002223290503, - "total_bits": 7950976, + "accuracy": 0.9953756900504231, + "total_bits": 23630336, "q_proj": { "group_size": { "5": 64, @@ -20311,8 +20311,8 @@ } }, { - "accuracy": 0.9911674838513136, - "total_bits": 8006912, + "accuracy": 0.9956478681415319, + "total_bits": 23872512, "q_proj": { "group_size": { "5": 32, @@ -20372,8 +20372,8 @@ } }, { - "accuracy": 0.9946174086071551, - "total_bits": 9727296, + "accuracy": 0.9975450842175633, + "total_bits": 29174016, "q_proj": { "group_size": { "6": 128, @@ -20433,8 +20433,8 @@ } }, { - "accuracy": 0.9950932073406875, - "total_bits": 9841920, + "accuracy": 0.9978302954696119, + "total_bits": 29377536, "q_proj": { "group_size": { "6": 32, @@ -20494,8 +20494,8 @@ } }, { - "accuracy": 0.9957236931659281, - "total_bits": 11201856, + "accuracy": 0.998314315569587, + "total_bits": 33433856, "q_proj": { "group_size": { "6": 128 @@ -20546,8 +20546,8 @@ } }, { - "accuracy": 0.9968434632755816, - "total_bits": 11607296, + "accuracy": 0.9988213798496872, + "total_bits": 34743296, "q_proj": { "group_size": { "6": 32 @@ -20598,8 +20598,8 @@ } }, { - "accuracy": 0.9988751251948997, - "total_bits": 14871872, + "accuracy": 0.9995475046744104, + "total_bits": 44443904, "q_proj": { "group_size": { "8": 128 @@ -20652,8 +20652,8 @@ ], "model.layers.10.mlp": [ { - "accuracy": 0.9586208239197731, - "total_bits": 29656128, + "accuracy": 0.9702269118279219, + "total_bits": 93392016, "gate_proj": { "group_size": { "3": 64, @@ -20704,8 +20704,8 @@ } }, { - "accuracy": 0.9605779349803925, - "total_bits": 30901312, + "accuracy": 0.9711166340857744, + "total_bits": 96259216, "gate_proj": { "group_size": { "3": 64, @@ -20756,8 +20756,8 @@ } }, { - "accuracy": 0.9657147899270058, - "total_bits": 34456000, + "accuracy": 0.9752719718962908, + "total_bits": 107863424, "gate_proj": { "group_size": { "3": 64, @@ -20805,8 +20805,8 @@ } }, { - "accuracy": 0.9673203937709332, - "total_bits": 38584768, + "accuracy": 0.9764827378094196, + "total_bits": 120937856, "gate_proj": { "group_size": { "3": 64, @@ -20854,8 +20854,8 @@ } }, { - "accuracy": 0.978621382266283, - "total_bits": 43342560, + "accuracy": 0.984815064817667, + "total_bits": 136162224, "gate_proj": { "group_size": { "4": 128, @@ -20906,8 +20906,8 @@ } }, { - "accuracy": 0.980769632384181, - "total_bits": 44242688, + "accuracy": 0.9861932778730989, + "total_bits": 139118208, "gate_proj": { "group_size": { "4": 32, @@ -20958,8 +20958,8 @@ } }, { - "accuracy": 0.9827583935111761, - "total_bits": 47520448, + "accuracy": 0.9877545088529587, + "total_bits": 149584176, "gate_proj": { "group_size": { "4": 32, @@ -21007,8 +21007,8 @@ } }, { - "accuracy": 0.9885018253698945, - "total_bits": 54270672, + "accuracy": 0.9920157203450799, + "total_bits": 170943088, "gate_proj": { "group_size": { "4": 128 @@ -21050,8 +21050,8 @@ } }, { - "accuracy": 0.9895953228697181, - "total_bits": 55072512, + "accuracy": 0.9927399610169232, + "total_bits": 173467264, "gate_proj": { "group_size": { "4": 32 @@ -21093,8 +21093,8 @@ } }, { - "accuracy": 0.9889660691842437, - "total_bits": 56187616, + "accuracy": 0.9921891749836504, + "total_bits": 176761776, "gate_proj": { "group_size": { "5": 128, @@ -21145,8 +21145,8 @@ } }, { - "accuracy": 0.9905442073941231, - "total_bits": 57087744, + "accuracy": 0.9932435471564531, + "total_bits": 179717760, "gate_proj": { "group_size": { "5": 32, @@ -21197,8 +21197,8 @@ } }, { - "accuracy": 0.9943201658315957, - "total_bits": 69032672, + "accuracy": 0.9959925631992519, + "total_bits": 217361328, "gate_proj": { "group_size": { "6": 128, @@ -21249,8 +21249,8 @@ } }, { - "accuracy": 0.9952527931891382, - "total_bits": 69932800, + "accuracy": 0.9966078596189618, + "total_bits": 220317312, "gate_proj": { "group_size": { "6": 32, @@ -21301,8 +21301,8 @@ } }, { - "accuracy": 0.9967377143912017, - "total_bits": 79551200, + "accuracy": 0.9977801421191543, + "total_bits": 250850224, "gate_proj": { "group_size": { "6": 128 @@ -21344,8 +21344,8 @@ } }, { - "accuracy": 0.9970799488946795, - "total_bits": 82937280, + "accuracy": 0.9979668704327196, + "total_bits": 262906272, "gate_proj": { "group_size": { "8": 128, @@ -21393,8 +21393,8 @@ } }, { - "accuracy": 0.9975337306968868, - "total_bits": 90277312, + "accuracy": 0.9983273103134707, + "total_bits": 286106016, "gate_proj": { "group_size": { "8": 128, @@ -21439,8 +21439,8 @@ } }, { - "accuracy": 0.9991097575984895, - "total_bits": 105219520, + "accuracy": 0.9994044461636804, + "total_bits": 331981216, "gate_proj": { "group_size": { "8": 128 @@ -21481,8 +21481,8 @@ ], "model.layers.11.self_attn": [ { - "accuracy": 0.9547368660569191, - "total_bits": 4051584, + "accuracy": 0.9776046220213175, + "total_bits": 12046848, "q_proj": { "group_size": { "3": 64, @@ -21545,8 +21545,8 @@ } }, { - "accuracy": 0.9567295871675014, - "total_bits": 4199040, + "accuracy": 0.9786577206104994, + "total_bits": 12325376, "q_proj": { "group_size": { "3": 64, @@ -21609,8 +21609,8 @@ } }, { - "accuracy": 0.9576077982783318, - "total_bits": 4293424, + "accuracy": 0.9812849536538124, + "total_bits": 12672960, "q_proj": { "group_size": { "3": 64, @@ -21673,8 +21673,8 @@ } }, { - "accuracy": 0.960216224193573, - "total_bits": 5070816, + "accuracy": 0.9865910233929753, + "total_bits": 15056256, "q_proj": { "group_size": { "3": 64, @@ -21737,8 +21737,8 @@ } }, { - "accuracy": 0.972936449572444, - "total_bits": 5958976, + "accuracy": 0.9889151817187667, + "total_bits": 17836288, "q_proj": { "group_size": { "4": 128, @@ -21801,8 +21801,8 @@ } }, { - "accuracy": 0.972521647810936, - "total_bits": 5962896, + "accuracy": 0.9890467831864953, + "total_bits": 17832768, "q_proj": { "group_size": { "4": 128, @@ -21865,8 +21865,8 @@ } }, { - "accuracy": 0.9792614672333002, - "total_bits": 7531840, + "accuracy": 0.9935071663931012, + "total_bits": 22423808, "q_proj": { "group_size": { "4": 128 @@ -21917,8 +21917,8 @@ } }, { - "accuracy": 0.9778816103935242, - "total_bits": 7535760, + "accuracy": 0.9937065038830042, + "total_bits": 22436672, "q_proj": { "group_size": { "4": 128 @@ -21969,8 +21969,8 @@ } }, { - "accuracy": 0.9803351312875748, - "total_bits": 7598368, + "accuracy": 0.9940308667719364, + "total_bits": 22623872, "q_proj": { "group_size": { "4": 64 @@ -22021,8 +22021,8 @@ } }, { - "accuracy": 0.979700593277812, - "total_bits": 7707904, + "accuracy": 0.9944512597285211, + "total_bits": 22946816, "q_proj": { "group_size": { "4": 32 @@ -22073,8 +22073,8 @@ } }, { - "accuracy": 0.9828558322042227, - "total_bits": 7797904, + "accuracy": 0.9943975452333689, + "total_bits": 23337792, "q_proj": { "group_size": { "5": 128, @@ -22137,8 +22137,8 @@ } }, { - "accuracy": 0.9833105728030205, - "total_bits": 7856416, + "accuracy": 0.9946797541342676, + "total_bits": 23303808, "q_proj": { "group_size": { "5": 64, @@ -22201,8 +22201,8 @@ } }, { - "accuracy": 0.9865790568292141, - "total_bits": 7950976, + "accuracy": 0.995271417312324, + "total_bits": 23630336, "q_proj": { "group_size": { "5": 64, @@ -22262,8 +22262,8 @@ } }, { - "accuracy": 0.9879224104806781, - "total_bits": 8006912, + "accuracy": 0.9956227042712271, + "total_bits": 23872512, "q_proj": { "group_size": { "5": 32, @@ -22323,8 +22323,8 @@ } }, { - "accuracy": 0.9912111917510629, - "total_bits": 9727296, + "accuracy": 0.9974854283500463, + "total_bits": 29174016, "q_proj": { "group_size": { "6": 128, @@ -22384,8 +22384,8 @@ } }, { - "accuracy": 0.9937804061919451, - "total_bits": 9841920, + "accuracy": 0.9977978523820639, + "total_bits": 29377536, "q_proj": { "group_size": { "6": 32, @@ -22445,8 +22445,8 @@ } }, { - "accuracy": 0.9936533416621387, - "total_bits": 11201856, + "accuracy": 0.9982931369449943, + "total_bits": 33433856, "q_proj": { "group_size": { "6": 128 @@ -22497,8 +22497,8 @@ } }, { - "accuracy": 0.9962142717558891, - "total_bits": 11607296, + "accuracy": 0.9988284043502063, + "total_bits": 34743296, "q_proj": { "group_size": { "6": 32 @@ -22549,8 +22549,8 @@ } }, { - "accuracy": 0.9981277505867183, - "total_bits": 14871872, + "accuracy": 0.9995489969733171, + "total_bits": 44443904, "q_proj": { "group_size": { "8": 128 @@ -22603,8 +22603,8 @@ ], "model.layers.11.mlp": [ { - "accuracy": 0.951299887150526, - "total_bits": 29656128, + "accuracy": 0.9718667417764664, + "total_bits": 93392016, "gate_proj": { "group_size": { "3": 64, @@ -22655,8 +22655,8 @@ } }, { - "accuracy": 0.95562943816185, - "total_bits": 30901312, + "accuracy": 0.9727223012596369, + "total_bits": 96259216, "gate_proj": { "group_size": { "3": 64, @@ -22707,8 +22707,8 @@ } }, { - "accuracy": 0.961256317794323, - "total_bits": 34456000, + "accuracy": 0.9763247519731522, + "total_bits": 107863424, "gate_proj": { "group_size": { "3": 64, @@ -22756,8 +22756,8 @@ } }, { - "accuracy": 0.9630678184330463, - "total_bits": 38584768, + "accuracy": 0.9774151276797056, + "total_bits": 120937856, "gate_proj": { "group_size": { "3": 64, @@ -22805,8 +22805,8 @@ } }, { - "accuracy": 0.9749839920550585, - "total_bits": 43342560, + "accuracy": 0.9856806537136436, + "total_bits": 136162224, "gate_proj": { "group_size": { "4": 128, @@ -22857,8 +22857,8 @@ } }, { - "accuracy": 0.9780963249504566, - "total_bits": 44242688, + "accuracy": 0.9870169907808304, + "total_bits": 139118208, "gate_proj": { "group_size": { "4": 32, @@ -22909,8 +22909,8 @@ } }, { - "accuracy": 0.980388605967164, - "total_bits": 47520448, + "accuracy": 0.9883980313315988, + "total_bits": 149584176, "gate_proj": { "group_size": { "4": 32, @@ -22958,8 +22958,8 @@ } }, { - "accuracy": 0.9864653367549181, - "total_bits": 54270672, + "accuracy": 0.9924592152237892, + "total_bits": 170943088, "gate_proj": { "group_size": { "4": 128 @@ -23001,8 +23001,8 @@ } }, { - "accuracy": 0.9876050408929586, - "total_bits": 55072512, + "accuracy": 0.9931318457238376, + "total_bits": 173467264, "gate_proj": { "group_size": { "4": 32 @@ -23044,8 +23044,8 @@ } }, { - "accuracy": 0.9867556942626834, - "total_bits": 56187616, + "accuracy": 0.9926311736926436, + "total_bits": 176761776, "gate_proj": { "group_size": { "5": 128, @@ -23096,8 +23096,8 @@ } }, { - "accuracy": 0.9890436576679349, - "total_bits": 57087744, + "accuracy": 0.993640061467886, + "total_bits": 179717760, "gate_proj": { "group_size": { "5": 32, @@ -23148,8 +23148,8 @@ } }, { - "accuracy": 0.9931194502860308, - "total_bits": 69032672, + "accuracy": 0.9962089876644313, + "total_bits": 217361328, "gate_proj": { "group_size": { "6": 128, @@ -23200,8 +23200,8 @@ } }, { - "accuracy": 0.9944707150571048, - "total_bits": 69932800, + "accuracy": 0.9968089594040066, + "total_bits": 220317312, "gate_proj": { "group_size": { "6": 32, @@ -23252,8 +23252,8 @@ } }, { - "accuracy": 0.9954788777977228, - "total_bits": 79551200, + "accuracy": 0.9978838765528053, + "total_bits": 250850224, "gate_proj": { "group_size": { "6": 128 @@ -23295,8 +23295,8 @@ } }, { - "accuracy": 0.9961503620725125, - "total_bits": 82937280, + "accuracy": 0.9980750273680314, + "total_bits": 262906272, "gate_proj": { "group_size": { "8": 128, @@ -23344,8 +23344,8 @@ } }, { - "accuracy": 0.9965920995455235, - "total_bits": 90277312, + "accuracy": 0.9983843646477908, + "total_bits": 286106016, "gate_proj": { "group_size": { "8": 128, @@ -23390,8 +23390,8 @@ } }, { - "accuracy": 0.998733103973791, - "total_bits": 105219520, + "accuracy": 0.9994237422361039, + "total_bits": 331981216, "gate_proj": { "group_size": { "8": 128 @@ -23432,8 +23432,8 @@ ], "model.layers.12.self_attn": [ { - "accuracy": 0.9617330878973007, - "total_bits": 4051584, + "accuracy": 0.978738447651267, + "total_bits": 12046848, "q_proj": { "group_size": { "3": 64, @@ -23496,8 +23496,8 @@ } }, { - "accuracy": 0.9654262773692608, - "total_bits": 4199040, + "accuracy": 0.9792955499142408, + "total_bits": 12325376, "q_proj": { "group_size": { "3": 64, @@ -23560,8 +23560,8 @@ } }, { - "accuracy": 0.96855678409338, - "total_bits": 4293424, + "accuracy": 0.9805421233177185, + "total_bits": 12672960, "q_proj": { "group_size": { "3": 64, @@ -23624,8 +23624,8 @@ } }, { - "accuracy": 0.9751867577433586, - "total_bits": 5070816, + "accuracy": 0.9851964935660362, + "total_bits": 15056256, "q_proj": { "group_size": { "3": 64, @@ -23688,8 +23688,8 @@ } }, { - "accuracy": 0.9809410944581032, - "total_bits": 5958976, + "accuracy": 0.9889361411333084, + "total_bits": 17836288, "q_proj": { "group_size": { "4": 128, @@ -23752,8 +23752,8 @@ } }, { - "accuracy": 0.981406319886446, - "total_bits": 5962896, + "accuracy": 0.9890276463702321, + "total_bits": 17832768, "q_proj": { "group_size": { "4": 128, @@ -23816,8 +23816,8 @@ } }, { - "accuracy": 0.9860723055899143, - "total_bits": 7531840, + "accuracy": 0.9930114774033427, + "total_bits": 22423808, "q_proj": { "group_size": { "4": 128 @@ -23868,8 +23868,8 @@ } }, { - "accuracy": 0.9866839628666639, - "total_bits": 7535760, + "accuracy": 0.9930966077372432, + "total_bits": 22436672, "q_proj": { "group_size": { "4": 128 @@ -23920,8 +23920,8 @@ } }, { - "accuracy": 0.9864359814673662, - "total_bits": 7598368, + "accuracy": 0.9936944833025336, + "total_bits": 22623872, "q_proj": { "group_size": { "4": 64 @@ -23972,8 +23972,8 @@ } }, { - "accuracy": 0.9869756614789367, - "total_bits": 7707904, + "accuracy": 0.9945216374471784, + "total_bits": 22946816, "q_proj": { "group_size": { "4": 32 @@ -24024,8 +24024,8 @@ } }, { - "accuracy": 0.9899861132726073, - "total_bits": 7797904, + "accuracy": 0.9945130385458469, + "total_bits": 23337792, "q_proj": { "group_size": { "5": 128, @@ -24088,8 +24088,8 @@ } }, { - "accuracy": 0.9905922180041671, - "total_bits": 7856416, + "accuracy": 0.9947595102712512, + "total_bits": 23303808, "q_proj": { "group_size": { "5": 64, @@ -24152,8 +24152,8 @@ } }, { - "accuracy": 0.9911313550546765, - "total_bits": 7950976, + "accuracy": 0.9950367962010205, + "total_bits": 23630336, "q_proj": { "group_size": { "5": 64, @@ -24213,8 +24213,8 @@ } }, { - "accuracy": 0.9911767523735762, - "total_bits": 8006912, + "accuracy": 0.9954659370705485, + "total_bits": 23872512, "q_proj": { "group_size": { "5": 32, @@ -24274,8 +24274,8 @@ } }, { - "accuracy": 0.9952182155102491, - "total_bits": 9727296, + "accuracy": 0.9973569277208298, + "total_bits": 29174016, "q_proj": { "group_size": { "6": 128, @@ -24335,8 +24335,8 @@ } }, { - "accuracy": 0.9958199453540146, - "total_bits": 9841920, + "accuracy": 0.997723177075386, + "total_bits": 29377536, "q_proj": { "group_size": { "6": 32, @@ -24396,8 +24396,8 @@ } }, { - "accuracy": 0.9961489604320377, - "total_bits": 11201856, + "accuracy": 0.9983315742574632, + "total_bits": 33433856, "q_proj": { "group_size": { "6": 128 @@ -24448,8 +24448,8 @@ } }, { - "accuracy": 0.9976966136600822, - "total_bits": 11607296, + "accuracy": 0.998701416188851, + "total_bits": 34743296, "q_proj": { "group_size": { "6": 32 @@ -24500,8 +24500,8 @@ } }, { - "accuracy": 0.998866674141027, - "total_bits": 14871872, + "accuracy": 0.9995390042895451, + "total_bits": 44443904, "q_proj": { "group_size": { "8": 128 @@ -24554,8 +24554,8 @@ ], "model.layers.12.mlp": [ { - "accuracy": 0.9576624184846878, - "total_bits": 29656128, + "accuracy": 0.9713144991546869, + "total_bits": 93392016, "gate_proj": { "group_size": { "3": 64, @@ -24606,8 +24606,8 @@ } }, { - "accuracy": 0.9595545157790184, - "total_bits": 30901312, + "accuracy": 0.9721826836466789, + "total_bits": 96259216, "gate_proj": { "group_size": { "3": 64, @@ -24658,8 +24658,8 @@ } }, { - "accuracy": 0.9651938080787659, - "total_bits": 34456000, + "accuracy": 0.9758257586508989, + "total_bits": 107863424, "gate_proj": { "group_size": { "3": 64, @@ -24707,8 +24707,8 @@ } }, { - "accuracy": 0.9668771326541901, - "total_bits": 38584768, + "accuracy": 0.9769103471189737, + "total_bits": 120937856, "gate_proj": { "group_size": { "3": 64, @@ -24756,8 +24756,8 @@ } }, { - "accuracy": 0.9780823178589344, - "total_bits": 43342560, + "accuracy": 0.9850145755335689, + "total_bits": 136162224, "gate_proj": { "group_size": { "4": 128, @@ -24808,8 +24808,8 @@ } }, { - "accuracy": 0.9802612587809563, - "total_bits": 44242688, + "accuracy": 0.9866860173642635, + "total_bits": 139118208, "gate_proj": { "group_size": { "4": 32, @@ -24860,8 +24860,8 @@ } }, { - "accuracy": 0.9824763238430023, - "total_bits": 47520448, + "accuracy": 0.9880683245137334, + "total_bits": 149584176, "gate_proj": { "group_size": { "4": 32, @@ -24909,8 +24909,8 @@ } }, { - "accuracy": 0.9883937854319811, - "total_bits": 54270672, + "accuracy": 0.9921933892183006, + "total_bits": 170943088, "gate_proj": { "group_size": { "4": 128 @@ -24952,8 +24952,8 @@ } }, { - "accuracy": 0.9894555266946554, - "total_bits": 55072512, + "accuracy": 0.9929307280108333, + "total_bits": 173467264, "gate_proj": { "group_size": { "4": 32 @@ -24995,8 +24995,8 @@ } }, { - "accuracy": 0.9886888358741999, - "total_bits": 56187616, + "accuracy": 0.992223936598748, + "total_bits": 176761776, "gate_proj": { "group_size": { "5": 128, @@ -25047,8 +25047,8 @@ } }, { - "accuracy": 0.990290317684412, - "total_bits": 57087744, + "accuracy": 0.9934494704939425, + "total_bits": 179717760, "gate_proj": { "group_size": { "5": 32, @@ -25099,8 +25099,8 @@ } }, { - "accuracy": 0.9941678480245173, - "total_bits": 69032672, + "accuracy": 0.9959858059883118, + "total_bits": 217361328, "gate_proj": { "group_size": { "6": 128, @@ -25151,8 +25151,8 @@ } }, { - "accuracy": 0.9951088149100542, - "total_bits": 69932800, + "accuracy": 0.9967025408986956, + "total_bits": 220317312, "gate_proj": { "group_size": { "6": 32, @@ -25203,8 +25203,8 @@ } }, { - "accuracy": 0.9967108147684485, - "total_bits": 79551200, + "accuracy": 0.9977564478758723, + "total_bits": 250850224, "gate_proj": { "group_size": { "6": 128 @@ -25246,8 +25246,8 @@ } }, { - "accuracy": 0.9969925913028419, - "total_bits": 82937280, + "accuracy": 0.9979549366980791, + "total_bits": 262906272, "gate_proj": { "group_size": { "8": 128, @@ -25295,8 +25295,8 @@ } }, { - "accuracy": 0.9974806155078113, - "total_bits": 90277312, + "accuracy": 0.9982591776642948, + "total_bits": 286106016, "gate_proj": { "group_size": { "8": 128, @@ -25341,8 +25341,8 @@ } }, { - "accuracy": 0.9991121572675183, - "total_bits": 105219520, + "accuracy": 0.9993936690152623, + "total_bits": 331981216, "gate_proj": { "group_size": { "8": 128 @@ -25383,8 +25383,8 @@ ], "model.layers.13.self_attn": [ { - "accuracy": 0.9414880238473415, - "total_bits": 4051584, + "accuracy": 0.9782510995864868, + "total_bits": 12046848, "q_proj": { "group_size": { "3": 64, @@ -25447,8 +25447,8 @@ } }, { - "accuracy": 0.961026307195425, - "total_bits": 4199040, + "accuracy": 0.9791547860950232, + "total_bits": 12325376, "q_proj": { "group_size": { "3": 64, @@ -25511,8 +25511,8 @@ } }, { - "accuracy": 0.964385811239481, - "total_bits": 4293424, + "accuracy": 0.9812339041382074, + "total_bits": 12672960, "q_proj": { "group_size": { "3": 64, @@ -25575,8 +25575,8 @@ } }, { - "accuracy": 0.9702844619750977, - "total_bits": 5070816, + "accuracy": 0.9870870588347316, + "total_bits": 15056256, "q_proj": { "group_size": { "3": 64, @@ -25639,8 +25639,8 @@ } }, { - "accuracy": 0.9782393369823694, - "total_bits": 5958976, + "accuracy": 0.9892066568136215, + "total_bits": 17836288, "q_proj": { "group_size": { "4": 128, @@ -25703,8 +25703,8 @@ } }, { - "accuracy": 0.9782495219260454, - "total_bits": 5962896, + "accuracy": 0.9893298866227269, + "total_bits": 17832768, "q_proj": { "group_size": { "4": 128, @@ -25767,8 +25767,8 @@ } }, { - "accuracy": 0.9834292307496071, - "total_bits": 7531840, + "accuracy": 0.9934507319703698, + "total_bits": 22423808, "q_proj": { "group_size": { "4": 128 @@ -25819,8 +25819,8 @@ } }, { - "accuracy": 0.9834407810121775, - "total_bits": 7535760, + "accuracy": 0.993580381385982, + "total_bits": 22436672, "q_proj": { "group_size": { "4": 128 @@ -25871,8 +25871,8 @@ } }, { - "accuracy": 0.9840813595801592, - "total_bits": 7598368, + "accuracy": 0.9941709265112877, + "total_bits": 22623872, "q_proj": { "group_size": { "4": 64 @@ -25923,8 +25923,8 @@ } }, { - "accuracy": 0.9857926936820149, - "total_bits": 7707904, + "accuracy": 0.9945287387818098, + "total_bits": 22946816, "q_proj": { "group_size": { "4": 32 @@ -25975,8 +25975,8 @@ } }, { - "accuracy": 0.9891220610588789, - "total_bits": 7797904, + "accuracy": 0.9946142886765301, + "total_bits": 23337792, "q_proj": { "group_size": { "5": 128, @@ -26039,8 +26039,8 @@ } }, { - "accuracy": 0.9898319151252508, - "total_bits": 7856416, + "accuracy": 0.9948870684020221, + "total_bits": 23303808, "q_proj": { "group_size": { "5": 64, @@ -26103,8 +26103,8 @@ } }, { - "accuracy": 0.9899416258558631, - "total_bits": 7950976, + "accuracy": 0.9952750075608492, + "total_bits": 23630336, "q_proj": { "group_size": { "5": 64, @@ -26164,8 +26164,8 @@ } }, { - "accuracy": 0.9909057850018144, - "total_bits": 8006912, + "accuracy": 0.9956046771258116, + "total_bits": 23872512, "q_proj": { "group_size": { "5": 32, @@ -26225,8 +26225,8 @@ } }, { - "accuracy": 0.994629823602736, - "total_bits": 9727296, + "accuracy": 0.9974695318378508, + "total_bits": 29174016, "q_proj": { "group_size": { "6": 128, @@ -26286,8 +26286,8 @@ } }, { - "accuracy": 0.9953972222283483, - "total_bits": 9841920, + "accuracy": 0.9978137018624693, + "total_bits": 29377536, "q_proj": { "group_size": { "6": 32, @@ -26347,8 +26347,8 @@ } }, { - "accuracy": 0.9957709261216223, - "total_bits": 11201856, + "accuracy": 0.9983204207383096, + "total_bits": 33433856, "q_proj": { "group_size": { "6": 128 @@ -26399,8 +26399,8 @@ } }, { - "accuracy": 0.9970666337758303, - "total_bits": 11607296, + "accuracy": 0.9988288069143891, + "total_bits": 34743296, "q_proj": { "group_size": { "6": 32 @@ -26451,8 +26451,8 @@ } }, { - "accuracy": 0.9986704710172489, - "total_bits": 14871872, + "accuracy": 0.9995485266845208, + "total_bits": 44443904, "q_proj": { "group_size": { "8": 128 @@ -26505,8 +26505,8 @@ ], "model.layers.13.mlp": [ { - "accuracy": 0.9527515731751919, - "total_bits": 29656128, + "accuracy": 0.9735497571527958, + "total_bits": 93392016, "gate_proj": { "group_size": { "3": 64, @@ -26557,8 +26557,8 @@ } }, { - "accuracy": 0.9548482969403267, - "total_bits": 30901312, + "accuracy": 0.9742809869349003, + "total_bits": 96259216, "gate_proj": { "group_size": { "3": 64, @@ -26609,8 +26609,8 @@ } }, { - "accuracy": 0.9611956216394901, - "total_bits": 34456000, + "accuracy": 0.9779920522123575, + "total_bits": 107863424, "gate_proj": { "group_size": { "3": 64, @@ -26658,8 +26658,8 @@ } }, { - "accuracy": 0.9632368087768555, - "total_bits": 38584768, + "accuracy": 0.9790999554097652, + "total_bits": 120937856, "gate_proj": { "group_size": { "3": 64, @@ -26707,8 +26707,8 @@ } }, { - "accuracy": 0.9753489270806313, - "total_bits": 43342560, + "accuracy": 0.9864724995568395, + "total_bits": 136162224, "gate_proj": { "group_size": { "4": 128, @@ -26759,8 +26759,8 @@ } }, { - "accuracy": 0.9778500832617283, - "total_bits": 44242688, + "accuracy": 0.9877228308469057, + "total_bits": 139118208, "gate_proj": { "group_size": { "4": 32, @@ -26811,8 +26811,8 @@ } }, { - "accuracy": 0.9803678076714277, - "total_bits": 47520448, + "accuracy": 0.9891943344846368, + "total_bits": 149584176, "gate_proj": { "group_size": { "4": 32, @@ -26860,8 +26860,8 @@ } }, { - "accuracy": 0.9868465326726437, - "total_bits": 54270672, + "accuracy": 0.9928748682141304, + "total_bits": 170943088, "gate_proj": { "group_size": { "4": 128 @@ -26903,8 +26903,8 @@ } }, { - "accuracy": 0.9881267454475164, - "total_bits": 55072512, + "accuracy": 0.9935275404714048, + "total_bits": 173467264, "gate_proj": { "group_size": { "4": 32 @@ -26946,8 +26946,8 @@ } }, { - "accuracy": 0.9872974148020148, - "total_bits": 56187616, + "accuracy": 0.9930614992044866, + "total_bits": 176761776, "gate_proj": { "group_size": { "5": 128, @@ -26998,8 +26998,8 @@ } }, { - "accuracy": 0.9891140712425113, - "total_bits": 57087744, + "accuracy": 0.9939904608763754, + "total_bits": 179717760, "gate_proj": { "group_size": { "5": 32, @@ -27050,8 +27050,8 @@ } }, { - "accuracy": 0.9934819508343935, - "total_bits": 69032672, + "accuracy": 0.9964456527959555, + "total_bits": 217361328, "gate_proj": { "group_size": { "6": 128, @@ -27102,8 +27102,8 @@ } }, { - "accuracy": 0.9945394792594016, - "total_bits": 69932800, + "accuracy": 0.9969920769799501, + "total_bits": 220317312, "gate_proj": { "group_size": { "6": 32, @@ -27154,8 +27154,8 @@ } }, { - "accuracy": 0.99629170098342, - "total_bits": 79551200, + "accuracy": 0.9980254040565342, + "total_bits": 250850224, "gate_proj": { "group_size": { "6": 128 @@ -27197,8 +27197,8 @@ } }, { - "accuracy": 0.9966537330765277, - "total_bits": 82937280, + "accuracy": 0.9982014346169308, + "total_bits": 262906272, "gate_proj": { "group_size": { "8": 128, @@ -27246,8 +27246,8 @@ } }, { - "accuracy": 0.9972294853068888, - "total_bits": 90277312, + "accuracy": 0.9985204569529742, + "total_bits": 286106016, "gate_proj": { "group_size": { "8": 128, @@ -27292,8 +27292,8 @@ } }, { - "accuracy": 0.9989982589613646, - "total_bits": 105219520, + "accuracy": 0.9994643470272422, + "total_bits": 331981216, "gate_proj": { "group_size": { "8": 128 @@ -27334,8 +27334,8 @@ ], "model.layers.14.self_attn": [ { - "accuracy": 0.9488061629235744, - "total_bits": 4051584, + "accuracy": 0.9758468102663755, + "total_bits": 12046848, "q_proj": { "group_size": { "3": 64, @@ -27398,8 +27398,8 @@ } }, { - "accuracy": 0.953201599419117, - "total_bits": 4199040, + "accuracy": 0.9768236298114061, + "total_bits": 12325376, "q_proj": { "group_size": { "3": 64, @@ -27462,8 +27462,8 @@ } }, { - "accuracy": 0.956532433629036, - "total_bits": 4293424, + "accuracy": 0.9785462897270918, + "total_bits": 12672960, "q_proj": { "group_size": { "3": 64, @@ -27526,8 +27526,8 @@ } }, { - "accuracy": 0.965791504830122, - "total_bits": 5070816, + "accuracy": 0.9828742723912001, + "total_bits": 15056256, "q_proj": { "group_size": { "3": 64, @@ -27590,8 +27590,8 @@ } }, { - "accuracy": 0.9728688318282366, - "total_bits": 5958976, + "accuracy": 0.9873956982046366, + "total_bits": 17836288, "q_proj": { "group_size": { "4": 128, @@ -27654,8 +27654,8 @@ } }, { - "accuracy": 0.973620880395174, - "total_bits": 5962896, + "accuracy": 0.9876556396484375, + "total_bits": 17832768, "q_proj": { "group_size": { "4": 128, @@ -27718,8 +27718,8 @@ } }, { - "accuracy": 0.9804780427366495, - "total_bits": 7531840, + "accuracy": 0.9912875406444073, + "total_bits": 22423808, "q_proj": { "group_size": { "4": 128 @@ -27770,8 +27770,8 @@ } }, { - "accuracy": 0.9812801815569401, - "total_bits": 7535760, + "accuracy": 0.9915633033961058, + "total_bits": 22436672, "q_proj": { "group_size": { "4": 128 @@ -27822,8 +27822,8 @@ } }, { - "accuracy": 0.98292126506567, - "total_bits": 7598368, + "accuracy": 0.9922241228632629, + "total_bits": 22623872, "q_proj": { "group_size": { "4": 64 @@ -27874,8 +27874,8 @@ } }, { - "accuracy": 0.9855556776747108, - "total_bits": 7707904, + "accuracy": 0.9927317681722343, + "total_bits": 22946816, "q_proj": { "group_size": { "4": 32 @@ -27926,8 +27926,8 @@ } }, { - "accuracy": 0.9874063953757286, - "total_bits": 7797904, + "accuracy": 0.9935175008140504, + "total_bits": 23337792, "q_proj": { "group_size": { "5": 128, @@ -27990,8 +27990,8 @@ } }, { - "accuracy": 0.9876387370750308, - "total_bits": 7856416, + "accuracy": 0.9938403367996216, + "total_bits": 23303808, "q_proj": { "group_size": { "5": 64, @@ -28054,8 +28054,8 @@ } }, { - "accuracy": 0.9884621100500226, - "total_bits": 7950976, + "accuracy": 0.994054920040071, + "total_bits": 23630336, "q_proj": { "group_size": { "5": 64, @@ -28115,8 +28115,8 @@ } }, { - "accuracy": 0.9892682675272226, - "total_bits": 8006912, + "accuracy": 0.9945044135674834, + "total_bits": 23872512, "q_proj": { "group_size": { "5": 32, @@ -28176,8 +28176,8 @@ } }, { - "accuracy": 0.9939091373234987, - "total_bits": 9727296, + "accuracy": 0.9967493514996022, + "total_bits": 29174016, "q_proj": { "group_size": { "6": 128, @@ -28237,8 +28237,8 @@ } }, { - "accuracy": 0.994295141659677, - "total_bits": 9841920, + "accuracy": 0.9972120495513082, + "total_bits": 29377536, "q_proj": { "group_size": { "6": 32, @@ -28298,8 +28298,8 @@ } }, { - "accuracy": 0.9954260787926614, - "total_bits": 11201856, + "accuracy": 0.9975586482323706, + "total_bits": 33433856, "q_proj": { "group_size": { "6": 128 @@ -28350,8 +28350,8 @@ } }, { - "accuracy": 0.9961859700269997, - "total_bits": 11607296, + "accuracy": 0.9981665723025799, + "total_bits": 34743296, "q_proj": { "group_size": { "6": 32 @@ -28402,8 +28402,8 @@ } }, { - "accuracy": 0.9987689683912322, - "total_bits": 14871872, + "accuracy": 0.9993740836507641, + "total_bits": 44443904, "q_proj": { "group_size": { "8": 128 @@ -28456,8 +28456,8 @@ ], "model.layers.14.mlp": [ { - "accuracy": 0.9515498876571655, - "total_bits": 29656128, + "accuracy": 0.9715385343879461, + "total_bits": 93392016, "gate_proj": { "group_size": { "3": 64, @@ -28508,8 +28508,8 @@ } }, { - "accuracy": 0.953888975083828, - "total_bits": 30901312, + "accuracy": 0.9724952317774296, + "total_bits": 96259216, "gate_proj": { "group_size": { "3": 64, @@ -28560,8 +28560,8 @@ } }, { - "accuracy": 0.9599342197179794, - "total_bits": 34456000, + "accuracy": 0.9765123184770346, + "total_bits": 107863424, "gate_proj": { "group_size": { "3": 64, @@ -28609,8 +28609,8 @@ } }, { - "accuracy": 0.9618732705712318, - "total_bits": 38584768, + "accuracy": 0.9776987582445145, + "total_bits": 120937856, "gate_proj": { "group_size": { "3": 64, @@ -28658,8 +28658,8 @@ } }, { - "accuracy": 0.9739084020256996, - "total_bits": 43342560, + "accuracy": 0.9853736218065023, + "total_bits": 136162224, "gate_proj": { "group_size": { "4": 128, @@ -28710,8 +28710,8 @@ } }, { - "accuracy": 0.9773633740842342, - "total_bits": 44242688, + "accuracy": 0.9867591708898544, + "total_bits": 139118208, "gate_proj": { "group_size": { "4": 32, @@ -28762,8 +28762,8 @@ } }, { - "accuracy": 0.9797431919723749, - "total_bits": 47520448, + "accuracy": 0.9882921762764454, + "total_bits": 149584176, "gate_proj": { "group_size": { "4": 32, @@ -28811,8 +28811,8 @@ } }, { - "accuracy": 0.9861893961206079, - "total_bits": 54270672, + "accuracy": 0.9922650698572397, + "total_bits": 170943088, "gate_proj": { "group_size": { "4": 128 @@ -28854,8 +28854,8 @@ } }, { - "accuracy": 0.9876388935372233, - "total_bits": 55072512, + "accuracy": 0.9929866786114872, + "total_bits": 173467264, "gate_proj": { "group_size": { "4": 32 @@ -28897,8 +28897,8 @@ } }, { - "accuracy": 0.9863216252997518, - "total_bits": 56187616, + "accuracy": 0.9924646578729153, + "total_bits": 176761776, "gate_proj": { "group_size": { "5": 128, @@ -28949,8 +28949,8 @@ } }, { - "accuracy": 0.9887453941628337, - "total_bits": 57087744, + "accuracy": 0.9935071007348597, + "total_bits": 179717760, "gate_proj": { "group_size": { "5": 32, @@ -29001,8 +29001,8 @@ } }, { - "accuracy": 0.9928995328955352, - "total_bits": 69032672, + "accuracy": 0.9961371910758317, + "total_bits": 217361328, "gate_proj": { "group_size": { "6": 128, @@ -29053,8 +29053,8 @@ } }, { - "accuracy": 0.9942919560708106, - "total_bits": 69932800, + "accuracy": 0.9967457184102386, + "total_bits": 220317312, "gate_proj": { "group_size": { "6": 32, @@ -29105,8 +29105,8 @@ } }, { - "accuracy": 0.9959470028989017, - "total_bits": 79551200, + "accuracy": 0.9978438755497336, + "total_bits": 250850224, "gate_proj": { "group_size": { "6": 128 @@ -29148,8 +29148,8 @@ } }, { - "accuracy": 0.9963231922592968, - "total_bits": 82937280, + "accuracy": 0.9980428912676871, + "total_bits": 262906272, "gate_proj": { "group_size": { "8": 128, @@ -29197,8 +29197,8 @@ } }, { - "accuracy": 0.9968288412783295, - "total_bits": 90277312, + "accuracy": 0.9983889113646001, + "total_bits": 286106016, "gate_proj": { "group_size": { "8": 128, @@ -29243,8 +29243,8 @@ } }, { - "accuracy": 0.9989001606591046, - "total_bits": 105219520, + "accuracy": 0.9994108344544657, + "total_bits": 331981216, "gate_proj": { "group_size": { "8": 128 @@ -29285,8 +29285,8 @@ ], "model.layers.15.self_attn": [ { - "accuracy": 0.9469720311462879, - "total_bits": 4051584, + "accuracy": 0.9775583744049072, + "total_bits": 12046848, "q_proj": { "group_size": { "3": 64, @@ -29349,8 +29349,8 @@ } }, { - "accuracy": 0.9545312970876694, - "total_bits": 4199040, + "accuracy": 0.978721346706152, + "total_bits": 12325376, "q_proj": { "group_size": { "3": 64, @@ -29413,8 +29413,8 @@ } }, { - "accuracy": 0.9571067430078983, - "total_bits": 4293424, + "accuracy": 0.9808999076485634, + "total_bits": 12672960, "q_proj": { "group_size": { "3": 64, @@ -29477,8 +29477,8 @@ } }, { - "accuracy": 0.9638321101665497, - "total_bits": 5070816, + "accuracy": 0.9869752693921328, + "total_bits": 15056256, "q_proj": { "group_size": { "3": 64, @@ -29541,8 +29541,8 @@ } }, { - "accuracy": 0.9745022673159838, - "total_bits": 5958976, + "accuracy": 0.9889641227200627, + "total_bits": 17836288, "q_proj": { "group_size": { "4": 128, @@ -29605,8 +29605,8 @@ } }, { - "accuracy": 0.9747154340147972, - "total_bits": 5962896, + "accuracy": 0.9890738800168037, + "total_bits": 17832768, "q_proj": { "group_size": { "4": 128, @@ -29669,8 +29669,8 @@ } }, { - "accuracy": 0.9810279291123152, - "total_bits": 7531840, + "accuracy": 0.9931364948861301, + "total_bits": 22423808, "q_proj": { "group_size": { "4": 128 @@ -29721,8 +29721,8 @@ } }, { - "accuracy": 0.9809613693505526, - "total_bits": 7535760, + "accuracy": 0.993351899087429, + "total_bits": 22436672, "q_proj": { "group_size": { "4": 128 @@ -29773,8 +29773,8 @@ } }, { - "accuracy": 0.9810642395168543, - "total_bits": 7598368, + "accuracy": 0.993881999514997, + "total_bits": 22623872, "q_proj": { "group_size": { "4": 64 @@ -29825,8 +29825,8 @@ } }, { - "accuracy": 0.985259672626853, - "total_bits": 7707904, + "accuracy": 0.9942388995550573, + "total_bits": 22946816, "q_proj": { "group_size": { "4": 32 @@ -29877,8 +29877,8 @@ } }, { - "accuracy": 0.9869902757927775, - "total_bits": 7797904, + "accuracy": 0.9943469949066639, + "total_bits": 23337792, "q_proj": { "group_size": { "5": 128, @@ -29941,8 +29941,8 @@ } }, { - "accuracy": 0.9889867110177875, - "total_bits": 7856416, + "accuracy": 0.9946556575596333, + "total_bits": 23303808, "q_proj": { "group_size": { "5": 64, @@ -30005,8 +30005,8 @@ } }, { - "accuracy": 0.989251553080976, - "total_bits": 7950976, + "accuracy": 0.9951433092355728, + "total_bits": 23630336, "q_proj": { "group_size": { "5": 64, @@ -30066,8 +30066,8 @@ } }, { - "accuracy": 0.9903627131134272, - "total_bits": 8006912, + "accuracy": 0.9954622220247984, + "total_bits": 23872512, "q_proj": { "group_size": { "5": 32, @@ -30127,8 +30127,8 @@ } }, { - "accuracy": 0.9938893462531269, - "total_bits": 9727296, + "accuracy": 0.9973680581897497, + "total_bits": 29174016, "q_proj": { "group_size": { "6": 128, @@ -30188,8 +30188,8 @@ } }, { - "accuracy": 0.9950756831094623, - "total_bits": 9841920, + "accuracy": 0.9976728889159858, + "total_bits": 29377536, "q_proj": { "group_size": { "6": 32, @@ -30249,8 +30249,8 @@ } }, { - "accuracy": 0.9952582567930222, - "total_bits": 11201856, + "accuracy": 0.9981526483315974, + "total_bits": 33433856, "q_proj": { "group_size": { "6": 128 @@ -30301,8 +30301,8 @@ } }, { - "accuracy": 0.9971242533065379, - "total_bits": 11607296, + "accuracy": 0.9986656652763486, + "total_bits": 34743296, "q_proj": { "group_size": { "6": 32 @@ -30353,8 +30353,8 @@ } }, { - "accuracy": 0.9988009865628555, - "total_bits": 14871872, + "accuracy": 0.9994497681036592, + "total_bits": 44443904, "q_proj": { "group_size": { "8": 128 @@ -30407,8 +30407,8 @@ ], "model.layers.15.mlp": [ { - "accuracy": 0.9447448067367077, - "total_bits": 29656128, + "accuracy": 0.9739448260515928, + "total_bits": 93392016, "gate_proj": { "group_size": { "3": 64, @@ -30459,8 +30459,8 @@ } }, { - "accuracy": 0.9478673562407494, - "total_bits": 30901312, + "accuracy": 0.97474598698318, + "total_bits": 96259216, "gate_proj": { "group_size": { "3": 64, @@ -30511,8 +30511,8 @@ } }, { - "accuracy": 0.954693891108036, - "total_bits": 34456000, + "accuracy": 0.9781367406249046, + "total_bits": 107863424, "gate_proj": { "group_size": { "3": 64, @@ -30560,8 +30560,8 @@ } }, { - "accuracy": 0.9569080695509911, - "total_bits": 38584768, + "accuracy": 0.9791515078395605, + "total_bits": 120937856, "gate_proj": { "group_size": { "3": 64, @@ -30609,8 +30609,8 @@ } }, { - "accuracy": 0.9714733213186264, - "total_bits": 43342560, + "accuracy": 0.9865438481792808, + "total_bits": 136162224, "gate_proj": { "group_size": { "4": 128, @@ -30661,8 +30661,8 @@ } }, { - "accuracy": 0.974591426551342, - "total_bits": 44242688, + "accuracy": 0.9878406189382076, + "total_bits": 139118208, "gate_proj": { "group_size": { "4": 32, @@ -30713,8 +30713,8 @@ } }, { - "accuracy": 0.9771509543061256, - "total_bits": 47520448, + "accuracy": 0.9891542578116059, + "total_bits": 149584176, "gate_proj": { "group_size": { "4": 32, @@ -30762,8 +30762,8 @@ } }, { - "accuracy": 0.9844310618937016, - "total_bits": 54270672, + "accuracy": 0.992917597759515, + "total_bits": 170943088, "gate_proj": { "group_size": { "4": 128 @@ -30805,8 +30805,8 @@ } }, { - "accuracy": 0.9860060941427946, - "total_bits": 55072512, + "accuracy": 0.9935614094138145, + "total_bits": 173467264, "gate_proj": { "group_size": { "4": 32 @@ -30848,8 +30848,8 @@ } }, { - "accuracy": 0.9852009760215878, - "total_bits": 56187616, + "accuracy": 0.9930817680433393, + "total_bits": 176761776, "gate_proj": { "group_size": { "5": 128, @@ -30900,8 +30900,8 @@ } }, { - "accuracy": 0.9874945851042867, - "total_bits": 57087744, + "accuracy": 0.9940295871347189, + "total_bits": 179717760, "gate_proj": { "group_size": { "5": 32, @@ -30952,8 +30952,8 @@ } }, { - "accuracy": 0.9923870316706598, - "total_bits": 69032672, + "accuracy": 0.9964420895557851, + "total_bits": 217361328, "gate_proj": { "group_size": { "6": 128, @@ -31004,8 +31004,8 @@ } }, { - "accuracy": 0.9936993620358407, - "total_bits": 69932800, + "accuracy": 0.9970052251592278, + "total_bits": 220317312, "gate_proj": { "group_size": { "6": 32, @@ -31056,8 +31056,8 @@ } }, { - "accuracy": 0.9955194243229926, - "total_bits": 79551200, + "accuracy": 0.9980168875772506, + "total_bits": 250850224, "gate_proj": { "group_size": { "6": 128 @@ -31099,8 +31099,8 @@ } }, { - "accuracy": 0.9960718997754157, - "total_bits": 82937280, + "accuracy": 0.9981989145744592, + "total_bits": 262906272, "gate_proj": { "group_size": { "8": 128, @@ -31148,8 +31148,8 @@ } }, { - "accuracy": 0.9966602816712111, - "total_bits": 90277312, + "accuracy": 0.9984951040241867, + "total_bits": 286106016, "gate_proj": { "group_size": { "8": 128, @@ -31194,8 +31194,8 @@ } }, { - "accuracy": 0.9987642507767305, - "total_bits": 105219520, + "accuracy": 0.9994575560558587, + "total_bits": 331981216, "gate_proj": { "group_size": { "8": 128 @@ -31236,8 +31236,8 @@ ], "model.layers.16.self_attn": [ { - "accuracy": 0.9391019158065319, - "total_bits": 4051584, + "accuracy": 0.9784835390746593, + "total_bits": 12046848, "q_proj": { "group_size": { "3": 64, @@ -31300,8 +31300,8 @@ } }, { - "accuracy": 0.9405354559421539, - "total_bits": 4199040, + "accuracy": 0.9790390394628048, + "total_bits": 12325376, "q_proj": { "group_size": { "3": 64, @@ -31364,8 +31364,8 @@ } }, { - "accuracy": 0.9455863051116467, - "total_bits": 4293424, + "accuracy": 0.9804434832185507, + "total_bits": 12672960, "q_proj": { "group_size": { "3": 64, @@ -31428,8 +31428,8 @@ } }, { - "accuracy": 0.9489683918654919, - "total_bits": 5070816, + "accuracy": 0.9851316250860691, + "total_bits": 15056256, "q_proj": { "group_size": { "3": 64, @@ -31492,8 +31492,8 @@ } }, { - "accuracy": 0.9613325372338295, - "total_bits": 5958976, + "accuracy": 0.9891424616798759, + "total_bits": 17836288, "q_proj": { "group_size": { "4": 128, @@ -31556,8 +31556,8 @@ } }, { - "accuracy": 0.9647579453885555, - "total_bits": 5962896, + "accuracy": 0.989252514205873, + "total_bits": 17832768, "q_proj": { "group_size": { "4": 128, @@ -31620,8 +31620,8 @@ } }, { - "accuracy": 0.9628359824419022, - "total_bits": 7531840, + "accuracy": 0.9935573409311473, + "total_bits": 22423808, "q_proj": { "group_size": { "4": 128 @@ -31672,8 +31672,8 @@ } }, { - "accuracy": 0.9716558326035738, - "total_bits": 7535760, + "accuracy": 0.9936770121566951, + "total_bits": 22436672, "q_proj": { "group_size": { "4": 128 @@ -31724,8 +31724,8 @@ } }, { - "accuracy": 0.9705669227987528, - "total_bits": 7598368, + "accuracy": 0.9939054911956191, + "total_bits": 22623872, "q_proj": { "group_size": { "4": 64 @@ -31776,8 +31776,8 @@ } }, { - "accuracy": 0.97020517103374, - "total_bits": 7707904, + "accuracy": 0.994319889228791, + "total_bits": 22946816, "q_proj": { "group_size": { "4": 32 @@ -31828,8 +31828,8 @@ } }, { - "accuracy": 0.9747034385800362, - "total_bits": 7797904, + "accuracy": 0.9945035143755376, + "total_bits": 23337792, "q_proj": { "group_size": { "5": 128, @@ -31892,8 +31892,8 @@ } }, { - "accuracy": 0.9752081613987684, - "total_bits": 7856416, + "accuracy": 0.9947340795770288, + "total_bits": 23303808, "q_proj": { "group_size": { "5": 64, @@ -31956,8 +31956,8 @@ } }, { - "accuracy": 0.9759049993008375, - "total_bits": 7950976, + "accuracy": 0.9950303053483367, + "total_bits": 23630336, "q_proj": { "group_size": { "5": 64, @@ -32017,8 +32017,8 @@ } }, { - "accuracy": 0.9782600942999125, - "total_bits": 8006912, + "accuracy": 0.9954373659566045, + "total_bits": 23872512, "q_proj": { "group_size": { "5": 32, @@ -32078,8 +32078,8 @@ } }, { - "accuracy": 0.9829973187297583, - "total_bits": 9727296, + "accuracy": 0.9973300914280117, + "total_bits": 29174016, "q_proj": { "group_size": { "6": 128, @@ -32139,8 +32139,8 @@ } }, { - "accuracy": 0.9863692400977015, - "total_bits": 9841920, + "accuracy": 0.9977092947810888, + "total_bits": 29377536, "q_proj": { "group_size": { "6": 32, @@ -32200,8 +32200,8 @@ } }, { - "accuracy": 0.9836499355733395, - "total_bits": 11201856, + "accuracy": 0.9983133522327989, + "total_bits": 33433856, "q_proj": { "group_size": { "6": 128 @@ -32252,8 +32252,8 @@ } }, { - "accuracy": 0.9953490453772247, - "total_bits": 11607296, + "accuracy": 0.9987422921694815, + "total_bits": 34743296, "q_proj": { "group_size": { "6": 32 @@ -32304,8 +32304,8 @@ } }, { - "accuracy": 0.9957951316609979, - "total_bits": 14871872, + "accuracy": 0.9995404232759029, + "total_bits": 44443904, "q_proj": { "group_size": { "8": 128 @@ -32358,8 +32358,8 @@ ], "model.layers.16.mlp": [ { - "accuracy": 0.9290044158697128, - "total_bits": 29656128, + "accuracy": 0.9722231440246105, + "total_bits": 93392016, "gate_proj": { "group_size": { "3": 64, @@ -32410,8 +32410,8 @@ } }, { - "accuracy": 0.933639258146286, - "total_bits": 30901312, + "accuracy": 0.9730795528739691, + "total_bits": 96259216, "gate_proj": { "group_size": { "3": 64, @@ -32462,8 +32462,8 @@ } }, { - "accuracy": 0.941821027547121, - "total_bits": 34456000, + "accuracy": 0.976656211540103, + "total_bits": 107863424, "gate_proj": { "group_size": { "3": 64, @@ -32511,8 +32511,8 @@ } }, { - "accuracy": 0.9447595626115799, - "total_bits": 38584768, + "accuracy": 0.9777514655143023, + "total_bits": 120937856, "gate_proj": { "group_size": { "3": 64, @@ -32560,8 +32560,8 @@ } }, { - "accuracy": 0.9621489383280277, - "total_bits": 43342560, + "accuracy": 0.9857162451371551, + "total_bits": 136162224, "gate_proj": { "group_size": { "4": 128, @@ -32612,8 +32612,8 @@ } }, { - "accuracy": 0.967151153832674, - "total_bits": 44242688, + "accuracy": 0.9870956875383854, + "total_bits": 139118208, "gate_proj": { "group_size": { "4": 32, @@ -32664,8 +32664,8 @@ } }, { - "accuracy": 0.9701167475432158, - "total_bits": 47520448, + "accuracy": 0.9884695559740067, + "total_bits": 149584176, "gate_proj": { "group_size": { "4": 32, @@ -32713,8 +32713,8 @@ } }, { - "accuracy": 0.9789236523211002, - "total_bits": 54270672, + "accuracy": 0.9924411550164223, + "total_bits": 170943088, "gate_proj": { "group_size": { "4": 128 @@ -32756,8 +32756,8 @@ } }, { - "accuracy": 0.9813529793173075, - "total_bits": 55072512, + "accuracy": 0.99312624335289, + "total_bits": 173467264, "gate_proj": { "group_size": { "4": 32 @@ -32799,8 +32799,8 @@ } }, { - "accuracy": 0.9800730235874653, - "total_bits": 56187616, + "accuracy": 0.9926464674063027, + "total_bits": 176761776, "gate_proj": { "group_size": { "5": 128, @@ -32851,8 +32851,8 @@ } }, { - "accuracy": 0.9836046379059553, - "total_bits": 57087744, + "accuracy": 0.9936684546992183, + "total_bits": 179717760, "gate_proj": { "group_size": { "5": 32, @@ -32903,8 +32903,8 @@ } }, { - "accuracy": 0.9896394982933998, - "total_bits": 69032672, + "accuracy": 0.9962192815728486, + "total_bits": 217361328, "gate_proj": { "group_size": { "6": 128, @@ -32955,8 +32955,8 @@ } }, { - "accuracy": 0.9916866598650813, - "total_bits": 69932800, + "accuracy": 0.996821609325707, + "total_bits": 220317312, "gate_proj": { "group_size": { "6": 32, @@ -33007,8 +33007,8 @@ } }, { - "accuracy": 0.9937013979069889, - "total_bits": 79551200, + "accuracy": 0.997874315828085, + "total_bits": 250850224, "gate_proj": { "group_size": { "6": 128 @@ -33050,8 +33050,8 @@ } }, { - "accuracy": 0.9944339841604233, - "total_bits": 82937280, + "accuracy": 0.9980859420029446, + "total_bits": 262906272, "gate_proj": { "group_size": { "8": 128, @@ -33099,8 +33099,8 @@ } }, { - "accuracy": 0.9951155297458172, - "total_bits": 90277312, + "accuracy": 0.9983942427206784, + "total_bits": 286106016, "gate_proj": { "group_size": { "8": 128, @@ -33145,8 +33145,8 @@ } }, { - "accuracy": 0.9982623364776373, - "total_bits": 105219520, + "accuracy": 0.9994219954824075, + "total_bits": 331981216, "gate_proj": { "group_size": { "8": 128 @@ -33187,8 +33187,8 @@ ], "model.layers.17.self_attn": [ { - "accuracy": 0.9462804310023785, - "total_bits": 4051584, + "accuracy": 0.9793892316520214, + "total_bits": 12046848, "q_proj": { "group_size": { "3": 64, @@ -33251,8 +33251,8 @@ } }, { - "accuracy": 0.9461779184639454, - "total_bits": 4199040, + "accuracy": 0.9799966439604759, + "total_bits": 12325376, "q_proj": { "group_size": { "3": 64, @@ -33315,8 +33315,8 @@ } }, { - "accuracy": 0.9509141594171524, - "total_bits": 4293424, + "accuracy": 0.982164666056633, + "total_bits": 12672960, "q_proj": { "group_size": { "3": 64, @@ -33379,8 +33379,8 @@ } }, { - "accuracy": 0.9533946216106415, - "total_bits": 5070816, + "accuracy": 0.986486753448844, + "total_bits": 15056256, "q_proj": { "group_size": { "3": 64, @@ -33443,8 +33443,8 @@ } }, { - "accuracy": 0.9674201421439648, - "total_bits": 5958976, + "accuracy": 0.9890371197834611, + "total_bits": 17836288, "q_proj": { "group_size": { "4": 128, @@ -33507,8 +33507,8 @@ } }, { - "accuracy": 0.9695307817310095, - "total_bits": 5962896, + "accuracy": 0.9891651310026646, + "total_bits": 17832768, "q_proj": { "group_size": { "4": 128, @@ -33571,8 +33571,8 @@ } }, { - "accuracy": 0.9805762898176908, - "total_bits": 7531840, + "accuracy": 0.9929002323187888, + "total_bits": 22423808, "q_proj": { "group_size": { "4": 128 @@ -33623,8 +33623,8 @@ } }, { - "accuracy": 0.9819745663553476, - "total_bits": 7535760, + "accuracy": 0.9931444427929819, + "total_bits": 22436672, "q_proj": { "group_size": { "4": 128 @@ -33675,8 +33675,8 @@ } }, { - "accuracy": 0.9843835923820734, - "total_bits": 7598368, + "accuracy": 0.9933849577791989, + "total_bits": 22623872, "q_proj": { "group_size": { "4": 64 @@ -33727,8 +33727,8 @@ } }, { - "accuracy": 0.9851059401407838, - "total_bits": 7707904, + "accuracy": 0.9946695752441883, + "total_bits": 22946816, "q_proj": { "group_size": { "4": 32 @@ -33779,8 +33779,8 @@ } }, { - "accuracy": 0.9834286123514175, - "total_bits": 7797904, + "accuracy": 0.9943312178365886, + "total_bits": 23337792, "q_proj": { "group_size": { "5": 128, @@ -33843,8 +33843,8 @@ } }, { - "accuracy": 0.9859553417190909, - "total_bits": 7856416, + "accuracy": 0.9949576170183718, + "total_bits": 23303808, "q_proj": { "group_size": { "5": 64, @@ -33907,8 +33907,8 @@ } }, { - "accuracy": 0.9889272712171078, - "total_bits": 7950976, + "accuracy": 0.9953497177921236, + "total_bits": 23630336, "q_proj": { "group_size": { "5": 64, @@ -33968,8 +33968,8 @@ } }, { - "accuracy": 0.9891826948150992, - "total_bits": 8006912, + "accuracy": 0.9956447859294713, + "total_bits": 23872512, "q_proj": { "group_size": { "5": 32, @@ -34029,8 +34029,8 @@ } }, { - "accuracy": 0.9935484500601888, - "total_bits": 9727296, + "accuracy": 0.9974364794325083, + "total_bits": 29174016, "q_proj": { "group_size": { "6": 128, @@ -34090,8 +34090,8 @@ } }, { - "accuracy": 0.9946266277693212, - "total_bits": 9841920, + "accuracy": 0.9978566558565944, + "total_bits": 29377536, "q_proj": { "group_size": { "6": 32, @@ -34151,8 +34151,8 @@ } }, { - "accuracy": 0.9947232198901474, - "total_bits": 11201856, + "accuracy": 0.9982390131335706, + "total_bits": 33433856, "q_proj": { "group_size": { "6": 128 @@ -34203,8 +34203,8 @@ } }, { - "accuracy": 0.997579267481342, - "total_bits": 11607296, + "accuracy": 0.9988534040749073, + "total_bits": 34743296, "q_proj": { "group_size": { "6": 32 @@ -34255,8 +34255,8 @@ } }, { - "accuracy": 0.9986451674485579, - "total_bits": 14871872, + "accuracy": 0.9995377856539562, + "total_bits": 44443904, "q_proj": { "group_size": { "8": 128 @@ -34309,8 +34309,8 @@ ], "model.layers.17.mlp": [ { - "accuracy": 0.9375919289886951, - "total_bits": 29656128, + "accuracy": 0.9702123552560806, + "total_bits": 93392016, "gate_proj": { "group_size": { "3": 64, @@ -34361,8 +34361,8 @@ } }, { - "accuracy": 0.9413831606507301, - "total_bits": 30901312, + "accuracy": 0.9711693860590458, + "total_bits": 96259216, "gate_proj": { "group_size": { "3": 64, @@ -34413,8 +34413,8 @@ } }, { - "accuracy": 0.9491118937730789, - "total_bits": 34456000, + "accuracy": 0.9751358013600111, + "total_bits": 107863424, "gate_proj": { "group_size": { "3": 64, @@ -34462,8 +34462,8 @@ } }, { - "accuracy": 0.9518261402845383, - "total_bits": 38584768, + "accuracy": 0.9763314817100763, + "total_bits": 120937856, "gate_proj": { "group_size": { "3": 64, @@ -34511,8 +34511,8 @@ } }, { - "accuracy": 0.9681604467332363, - "total_bits": 43342560, + "accuracy": 0.9845545785501599, + "total_bits": 136162224, "gate_proj": { "group_size": { "4": 128, @@ -34563,8 +34563,8 @@ } }, { - "accuracy": 0.9711183942854404, - "total_bits": 44242688, + "accuracy": 0.9861578531563282, + "total_bits": 139118208, "gate_proj": { "group_size": { "4": 32, @@ -34615,8 +34615,8 @@ } }, { - "accuracy": 0.9743590094149113, - "total_bits": 47520448, + "accuracy": 0.9877064544707537, + "total_bits": 149584176, "gate_proj": { "group_size": { "4": 32, @@ -34664,8 +34664,8 @@ } }, { - "accuracy": 0.9824552200734615, - "total_bits": 54270672, + "accuracy": 0.9919054042547941, + "total_bits": 170943088, "gate_proj": { "group_size": { "4": 128 @@ -34707,8 +34707,8 @@ } }, { - "accuracy": 0.9843091554939747, - "total_bits": 55072512, + "accuracy": 0.9926247703842819, + "total_bits": 173467264, "gate_proj": { "group_size": { "4": 32 @@ -34750,8 +34750,8 @@ } }, { - "accuracy": 0.9836191665381193, - "total_bits": 56187616, + "accuracy": 0.9920484172180295, + "total_bits": 176761776, "gate_proj": { "group_size": { "5": 128, @@ -34802,8 +34802,8 @@ } }, { - "accuracy": 0.9858829695731401, - "total_bits": 57087744, + "accuracy": 0.9932174878194928, + "total_bits": 179717760, "gate_proj": { "group_size": { "5": 32, @@ -34854,8 +34854,8 @@ } }, { - "accuracy": 0.9916049540042877, - "total_bits": 69032672, + "accuracy": 0.9959214883856475, + "total_bits": 217361328, "gate_proj": { "group_size": { "6": 128, @@ -34906,8 +34906,8 @@ } }, { - "accuracy": 0.9929212196730077, - "total_bits": 69932800, + "accuracy": 0.9965897465590388, + "total_bits": 220317312, "gate_proj": { "group_size": { "6": 32, @@ -34958,8 +34958,8 @@ } }, { - "accuracy": 0.9951141527853906, - "total_bits": 79551200, + "accuracy": 0.9977103781420738, + "total_bits": 250850224, "gate_proj": { "group_size": { "6": 128 @@ -35001,8 +35001,8 @@ } }, { - "accuracy": 0.9956942144781351, - "total_bits": 82937280, + "accuracy": 0.9979383677709848, + "total_bits": 262906272, "gate_proj": { "group_size": { "8": 128, @@ -35050,8 +35050,8 @@ } }, { - "accuracy": 0.9964349493384361, - "total_bits": 90277312, + "accuracy": 0.9982791322981939, + "total_bits": 286106016, "gate_proj": { "group_size": { "8": 128, @@ -35096,8 +35096,8 @@ } }, { - "accuracy": 0.9986615852685645, - "total_bits": 105219520, + "accuracy": 0.9993800488300622, + "total_bits": 331981216, "gate_proj": { "group_size": { "8": 128 @@ -35138,8 +35138,8 @@ ], "model.layers.18.self_attn": [ { - "accuracy": 0.9448805749416351, - "total_bits": 4051584, + "accuracy": 0.9761136621236801, + "total_bits": 12046848, "q_proj": { "group_size": { "3": 64, @@ -35202,8 +35202,8 @@ } }, { - "accuracy": 0.9464968964457512, - "total_bits": 4199040, + "accuracy": 0.9770606607198715, + "total_bits": 12325376, "q_proj": { "group_size": { "3": 64, @@ -35266,8 +35266,8 @@ } }, { - "accuracy": 0.9480113834142685, - "total_bits": 4293424, + "accuracy": 0.9795316681265831, + "total_bits": 12672960, "q_proj": { "group_size": { "3": 64, @@ -35330,8 +35330,8 @@ } }, { - "accuracy": 0.95137969404459, - "total_bits": 5070816, + "accuracy": 0.9853478269651532, + "total_bits": 15056256, "q_proj": { "group_size": { "3": 64, @@ -35394,8 +35394,8 @@ } }, { - "accuracy": 0.960879884660244, - "total_bits": 5958976, + "accuracy": 0.9882599031552672, + "total_bits": 17836288, "q_proj": { "group_size": { "4": 128, @@ -35458,8 +35458,8 @@ } }, { - "accuracy": 0.9624910429120064, - "total_bits": 5962896, + "accuracy": 0.9884275253862143, + "total_bits": 17832768, "q_proj": { "group_size": { "4": 128, @@ -35522,8 +35522,8 @@ } }, { - "accuracy": 0.9768306463956833, - "total_bits": 7531840, + "accuracy": 0.9926980817690492, + "total_bits": 22423808, "q_proj": { "group_size": { "4": 128 @@ -35574,8 +35574,8 @@ } }, { - "accuracy": 0.9778400100767612, - "total_bits": 7535760, + "accuracy": 0.9929794371128082, + "total_bits": 22436672, "q_proj": { "group_size": { "4": 128 @@ -35626,8 +35626,8 @@ } }, { - "accuracy": 0.9811704289168119, - "total_bits": 7598368, + "accuracy": 0.9934700899757445, + "total_bits": 22623872, "q_proj": { "group_size": { "4": 64 @@ -35678,8 +35678,8 @@ } }, { - "accuracy": 0.9818652179092169, - "total_bits": 7707904, + "accuracy": 0.9935703184455633, + "total_bits": 22946816, "q_proj": { "group_size": { "4": 32 @@ -35730,8 +35730,8 @@ } }, { - "accuracy": 0.9793904665857553, - "total_bits": 7797904, + "accuracy": 0.9941640924662352, + "total_bits": 23337792, "q_proj": { "group_size": { "5": 128, @@ -35794,8 +35794,8 @@ } }, { - "accuracy": 0.981003075838089, - "total_bits": 7856416, + "accuracy": 0.9944160711020231, + "total_bits": 23303808, "q_proj": { "group_size": { "5": 64, @@ -35858,8 +35858,8 @@ } }, { - "accuracy": 0.9839440882205963, - "total_bits": 7950976, + "accuracy": 0.994860075879842, + "total_bits": 23630336, "q_proj": { "group_size": { "5": 64, @@ -35919,8 +35919,8 @@ } }, { - "accuracy": 0.9853049078956246, - "total_bits": 8006912, + "accuracy": 0.9952407698146999, + "total_bits": 23872512, "q_proj": { "group_size": { "5": 32, @@ -35980,8 +35980,8 @@ } }, { - "accuracy": 0.9890499887987971, - "total_bits": 9727296, + "accuracy": 0.9972031733486801, + "total_bits": 29174016, "q_proj": { "group_size": { "6": 128, @@ -36041,8 +36041,8 @@ } }, { - "accuracy": 0.9929763246327639, - "total_bits": 9841920, + "accuracy": 0.9976053007412702, + "total_bits": 29377536, "q_proj": { "group_size": { "6": 32, @@ -36102,8 +36102,8 @@ } }, { - "accuracy": 0.9935302506200969, - "total_bits": 11201856, + "accuracy": 0.9980533812195063, + "total_bits": 33433856, "q_proj": { "group_size": { "6": 128 @@ -36154,8 +36154,8 @@ } }, { - "accuracy": 0.9970566479023546, - "total_bits": 11607296, + "accuracy": 0.9986692265374586, + "total_bits": 34743296, "q_proj": { "group_size": { "6": 32 @@ -36206,8 +36206,8 @@ } }, { - "accuracy": 0.9984088414348662, - "total_bits": 14871872, + "accuracy": 0.999480586964637, + "total_bits": 44443904, "q_proj": { "group_size": { "8": 128 @@ -36260,8 +36260,8 @@ ], "model.layers.18.mlp": [ { - "accuracy": 0.935174860060215, - "total_bits": 29656128, + "accuracy": 0.9668669477105141, + "total_bits": 93392016, "gate_proj": { "group_size": { "3": 64, @@ -36312,8 +36312,8 @@ } }, { - "accuracy": 0.9378093667328358, - "total_bits": 30901312, + "accuracy": 0.9679712615907192, + "total_bits": 96259216, "gate_proj": { "group_size": { "3": 64, @@ -36364,8 +36364,8 @@ } }, { - "accuracy": 0.9457938559353352, - "total_bits": 34456000, + "accuracy": 0.97240050137043, + "total_bits": 107863424, "gate_proj": { "group_size": { "3": 64, @@ -36413,8 +36413,8 @@ } }, { - "accuracy": 0.9484028145670891, - "total_bits": 38584768, + "accuracy": 0.9737879261374474, + "total_bits": 120937856, "gate_proj": { "group_size": { "3": 64, @@ -36462,8 +36462,8 @@ } }, { - "accuracy": 0.966415673494339, - "total_bits": 43342560, + "accuracy": 0.982800368219614, + "total_bits": 136162224, "gate_proj": { "group_size": { "4": 128, @@ -36514,8 +36514,8 @@ } }, { - "accuracy": 0.9695028308779001, - "total_bits": 44242688, + "accuracy": 0.9845740580931306, + "total_bits": 139118208, "gate_proj": { "group_size": { "4": 32, @@ -36566,8 +36566,8 @@ } }, { - "accuracy": 0.9728735014796257, - "total_bits": 47520448, + "accuracy": 0.9863665476441383, + "total_bits": 149584176, "gate_proj": { "group_size": { "4": 32, @@ -36615,8 +36615,8 @@ } }, { - "accuracy": 0.982090113684535, - "total_bits": 54270672, + "accuracy": 0.9908648300915956, + "total_bits": 170943088, "gate_proj": { "group_size": { "4": 128 @@ -36658,8 +36658,8 @@ } }, { - "accuracy": 0.9838197566568851, - "total_bits": 55072512, + "accuracy": 0.9917558152228594, + "total_bits": 173467264, "gate_proj": { "group_size": { "4": 32 @@ -36701,8 +36701,8 @@ } }, { - "accuracy": 0.9826987367123365, - "total_bits": 56187616, + "accuracy": 0.9911370426416397, + "total_bits": 176761776, "gate_proj": { "group_size": { "5": 128, @@ -36753,8 +36753,8 @@ } }, { - "accuracy": 0.9850574918091297, - "total_bits": 57087744, + "accuracy": 0.9924385696649551, + "total_bits": 179717760, "gate_proj": { "group_size": { "5": 32, @@ -36805,8 +36805,8 @@ } }, { - "accuracy": 0.9911482958123088, - "total_bits": 69032672, + "accuracy": 0.9954650518484414, + "total_bits": 217361328, "gate_proj": { "group_size": { "6": 128, @@ -36857,8 +36857,8 @@ } }, { - "accuracy": 0.9925176077522337, - "total_bits": 69932800, + "accuracy": 0.9962087315507233, + "total_bits": 220317312, "gate_proj": { "group_size": { "6": 32, @@ -36909,8 +36909,8 @@ } }, { - "accuracy": 0.99503427837044, - "total_bits": 79551200, + "accuracy": 0.9974413991440088, + "total_bits": 250850224, "gate_proj": { "group_size": { "6": 128 @@ -36952,8 +36952,8 @@ } }, { - "accuracy": 0.9954646187834442, - "total_bits": 82937280, + "accuracy": 0.9977091487962753, + "total_bits": 262906272, "gate_proj": { "group_size": { "8": 128, @@ -37001,8 +37001,8 @@ } }, { - "accuracy": 0.996174419298768, - "total_bits": 90277312, + "accuracy": 0.998112715431489, + "total_bits": 286106016, "gate_proj": { "group_size": { "8": 128, @@ -37047,8 +37047,8 @@ } }, { - "accuracy": 0.9986507809953764, - "total_bits": 105219520, + "accuracy": 0.9993042233400047, + "total_bits": 331981216, "gate_proj": { "group_size": { "8": 128 @@ -37089,8 +37089,8 @@ ], "model.layers.19.self_attn": [ { - "accuracy": 0.9427519664168358, - "total_bits": 4051584, + "accuracy": 0.9712088853120804, + "total_bits": 12046848, "q_proj": { "group_size": { "3": 64, @@ -37153,8 +37153,8 @@ } }, { - "accuracy": 0.9482091292738914, - "total_bits": 4199040, + "accuracy": 0.9724705368280411, + "total_bits": 12325376, "q_proj": { "group_size": { "3": 64, @@ -37217,8 +37217,8 @@ } }, { - "accuracy": 0.9502298906445503, - "total_bits": 4293424, + "accuracy": 0.9762942474335432, + "total_bits": 12672960, "q_proj": { "group_size": { "3": 64, @@ -37281,8 +37281,8 @@ } }, { - "accuracy": 0.9552766121923923, - "total_bits": 5070816, + "accuracy": 0.9815585892647505, + "total_bits": 15056256, "q_proj": { "group_size": { "3": 64, @@ -37345,8 +37345,8 @@ } }, { - "accuracy": 0.9695384744554758, - "total_bits": 5958976, + "accuracy": 0.9844543607905507, + "total_bits": 17836288, "q_proj": { "group_size": { "4": 128, @@ -37409,8 +37409,8 @@ } }, { - "accuracy": 0.9699725229293108, - "total_bits": 5962896, + "accuracy": 0.9852002300322056, + "total_bits": 17832768, "q_proj": { "group_size": { "4": 128, @@ -37473,8 +37473,8 @@ } }, { - "accuracy": 0.9822455178946257, - "total_bits": 7531840, + "accuracy": 0.9886356592178345, + "total_bits": 22423808, "q_proj": { "group_size": { "4": 128 @@ -37525,8 +37525,8 @@ } }, { - "accuracy": 0.9825639948248863, - "total_bits": 7535760, + "accuracy": 0.9898150656372309, + "total_bits": 22436672, "q_proj": { "group_size": { "4": 128 @@ -37577,8 +37577,8 @@ } }, { - "accuracy": 0.9836442414671183, - "total_bits": 7598368, + "accuracy": 0.9911776836961508, + "total_bits": 22623872, "q_proj": { "group_size": { "4": 64 @@ -37629,8 +37629,8 @@ } }, { - "accuracy": 0.9859087467193604, - "total_bits": 7707904, + "accuracy": 0.9913440952077508, + "total_bits": 22946816, "q_proj": { "group_size": { "4": 32 @@ -37681,8 +37681,8 @@ } }, { - "accuracy": 0.9839414302259684, - "total_bits": 7797904, + "accuracy": 0.9922102373093367, + "total_bits": 23337792, "q_proj": { "group_size": { "5": 128, @@ -37745,8 +37745,8 @@ } }, { - "accuracy": 0.9858170533552766, - "total_bits": 7856416, + "accuracy": 0.9925371371209621, + "total_bits": 23303808, "q_proj": { "group_size": { "5": 64, @@ -37809,8 +37809,8 @@ } }, { - "accuracy": 0.9876724136993289, - "total_bits": 7950976, + "accuracy": 0.9930709693580866, + "total_bits": 23630336, "q_proj": { "group_size": { "5": 64, @@ -37870,8 +37870,8 @@ } }, { - "accuracy": 0.990105714648962, - "total_bits": 8006912, + "accuracy": 0.9933527163229883, + "total_bits": 23872512, "q_proj": { "group_size": { "5": 32, @@ -37931,8 +37931,8 @@ } }, { - "accuracy": 0.9941023834981024, - "total_bits": 9727296, + "accuracy": 0.9961195737123489, + "total_bits": 29174016, "q_proj": { "group_size": { "6": 128, @@ -37992,8 +37992,8 @@ } }, { - "accuracy": 0.9948186413384974, - "total_bits": 9841920, + "accuracy": 0.9965932390186936, + "total_bits": 29377536, "q_proj": { "group_size": { "6": 32, @@ -38053,8 +38053,8 @@ } }, { - "accuracy": 0.9953572638332844, - "total_bits": 11201856, + "accuracy": 0.9967453263234347, + "total_bits": 33433856, "q_proj": { "group_size": { "6": 128 @@ -38105,8 +38105,8 @@ } }, { - "accuracy": 0.9974014775361866, - "total_bits": 11607296, + "accuracy": 0.9982008617371321, + "total_bits": 34743296, "q_proj": { "group_size": { "6": 32 @@ -38157,8 +38157,8 @@ } }, { - "accuracy": 0.9988005182240158, - "total_bits": 14871872, + "accuracy": 0.9991661829408258, + "total_bits": 44443904, "q_proj": { "group_size": { "8": 128 @@ -38211,8 +38211,8 @@ ], "model.layers.19.mlp": [ { - "accuracy": 0.9187815487384796, - "total_bits": 29656128, + "accuracy": 0.9562346860766411, + "total_bits": 93392016, "gate_proj": { "group_size": { "3": 64, @@ -38263,8 +38263,8 @@ } }, { - "accuracy": 0.9227375388145447, - "total_bits": 30901312, + "accuracy": 0.958115067332983, + "total_bits": 96259216, "gate_proj": { "group_size": { "3": 64, @@ -38315,8 +38315,8 @@ } }, { - "accuracy": 0.9315413534641266, - "total_bits": 34456000, + "accuracy": 0.963969275355339, + "total_bits": 107863424, "gate_proj": { "group_size": { "3": 64, @@ -38364,8 +38364,8 @@ } }, { - "accuracy": 0.9346312507987022, - "total_bits": 38584768, + "accuracy": 0.9657996296882629, + "total_bits": 120937856, "gate_proj": { "group_size": { "3": 64, @@ -38413,8 +38413,8 @@ } }, { - "accuracy": 0.9568981416523457, - "total_bits": 43342560, + "accuracy": 0.9774100985378027, + "total_bits": 136162224, "gate_proj": { "group_size": { "4": 128, @@ -38465,8 +38465,8 @@ } }, { - "accuracy": 0.9620590284466743, - "total_bits": 44242688, + "accuracy": 0.9797655288130045, + "total_bits": 139118208, "gate_proj": { "group_size": { "4": 32, @@ -38517,8 +38517,8 @@ } }, { - "accuracy": 0.9657231494784355, - "total_bits": 47520448, + "accuracy": 0.9819032084196806, + "total_bits": 149584176, "gate_proj": { "group_size": { "4": 32, @@ -38566,8 +38566,8 @@ } }, { - "accuracy": 0.976843198761344, - "total_bits": 54270672, + "accuracy": 0.9877016162499785, + "total_bits": 170943088, "gate_proj": { "group_size": { "4": 128 @@ -38609,8 +38609,8 @@ } }, { - "accuracy": 0.979301342740655, - "total_bits": 55072512, + "accuracy": 0.9889264507219195, + "total_bits": 173467264, "gate_proj": { "group_size": { "4": 32 @@ -38652,8 +38652,8 @@ } }, { - "accuracy": 0.9774577524513006, - "total_bits": 56187616, + "accuracy": 0.9881741367280483, + "total_bits": 176761776, "gate_proj": { "group_size": { "5": 128, @@ -38704,8 +38704,8 @@ } }, { - "accuracy": 0.9812527932226658, - "total_bits": 57087744, + "accuracy": 0.9899736940860748, + "total_bits": 179717760, "gate_proj": { "group_size": { "5": 32, @@ -38756,8 +38756,8 @@ } }, { - "accuracy": 0.9882521778345108, - "total_bits": 69032672, + "accuracy": 0.9938806081190705, + "total_bits": 217361328, "gate_proj": { "group_size": { "6": 128, @@ -38808,8 +38808,8 @@ } }, { - "accuracy": 0.9905463308095932, - "total_bits": 69932800, + "accuracy": 0.9949339954182506, + "total_bits": 220317312, "gate_proj": { "group_size": { "6": 32, @@ -38860,8 +38860,8 @@ } }, { - "accuracy": 0.9932587002404034, - "total_bits": 79551200, + "accuracy": 0.9964795028790832, + "total_bits": 250850224, "gate_proj": { "group_size": { "6": 128 @@ -38903,8 +38903,8 @@ } }, { - "accuracy": 0.9939715340733528, - "total_bits": 82937280, + "accuracy": 0.9968812663573772, + "total_bits": 262906272, "gate_proj": { "group_size": { "8": 128, @@ -38952,8 +38952,8 @@ } }, { - "accuracy": 0.9947369359433651, - "total_bits": 90277312, + "accuracy": 0.9973643801640719, + "total_bits": 286106016, "gate_proj": { "group_size": { "8": 128, @@ -38998,8 +38998,8 @@ } }, { - "accuracy": 0.9981496095424518, - "total_bits": 105219520, + "accuracy": 0.9990313068847172, + "total_bits": 331981216, "gate_proj": { "group_size": { "8": 128 @@ -39040,8 +39040,8 @@ ], "model.layers.20.self_attn": [ { - "accuracy": 0.9103880822658539, - "total_bits": 4051584, + "accuracy": 0.9649563394486904, + "total_bits": 12046848, "q_proj": { "group_size": { "3": 64, @@ -39104,8 +39104,8 @@ } }, { - "accuracy": 0.9254033416509628, - "total_bits": 4199040, + "accuracy": 0.966551773250103, + "total_bits": 12325376, "q_proj": { "group_size": { "3": 64, @@ -39168,8 +39168,8 @@ } }, { - "accuracy": 0.9280519410967827, - "total_bits": 4293424, + "accuracy": 0.9698956795036793, + "total_bits": 12672960, "q_proj": { "group_size": { "3": 64, @@ -39232,8 +39232,8 @@ } }, { - "accuracy": 0.9312246590852737, - "total_bits": 5070816, + "accuracy": 0.9772035572677851, + "total_bits": 15056256, "q_proj": { "group_size": { "3": 64, @@ -39296,8 +39296,8 @@ } }, { - "accuracy": 0.9454923719167709, - "total_bits": 5958976, + "accuracy": 0.9824510831385851, + "total_bits": 17836288, "q_proj": { "group_size": { "4": 128, @@ -39360,8 +39360,8 @@ } }, { - "accuracy": 0.9468598291277885, - "total_bits": 5962896, + "accuracy": 0.9826417490839958, + "total_bits": 17832768, "q_proj": { "group_size": { "4": 128, @@ -39424,8 +39424,8 @@ } }, { - "accuracy": 0.9637245163321495, - "total_bits": 7531840, + "accuracy": 0.989250972867012, + "total_bits": 22423808, "q_proj": { "group_size": { "4": 128 @@ -39476,8 +39476,8 @@ } }, { - "accuracy": 0.9670483395457268, - "total_bits": 7535760, + "accuracy": 0.989658435806632, + "total_bits": 22436672, "q_proj": { "group_size": { "4": 128 @@ -39528,8 +39528,8 @@ } }, { - "accuracy": 0.9731248281896114, - "total_bits": 7598368, + "accuracy": 0.9906487669795752, + "total_bits": 22623872, "q_proj": { "group_size": { "4": 64 @@ -39580,8 +39580,8 @@ } }, { - "accuracy": 0.9689498357474804, - "total_bits": 7707904, + "accuracy": 0.991015886887908, + "total_bits": 22946816, "q_proj": { "group_size": { "4": 32 @@ -39632,8 +39632,8 @@ } }, { - "accuracy": 0.9695971701294184, - "total_bits": 7797904, + "accuracy": 0.9910590630024672, + "total_bits": 23337792, "q_proj": { "group_size": { "5": 128, @@ -39696,8 +39696,8 @@ } }, { - "accuracy": 0.9761692713946104, - "total_bits": 7856416, + "accuracy": 0.9917043996974826, + "total_bits": 23303808, "q_proj": { "group_size": { "5": 64, @@ -39760,8 +39760,8 @@ } }, { - "accuracy": 0.9796416200697422, - "total_bits": 7950976, + "accuracy": 0.9924992783926427, + "total_bits": 23630336, "q_proj": { "group_size": { "5": 64, @@ -39821,8 +39821,8 @@ } }, { - "accuracy": 0.9810198377817869, - "total_bits": 8006912, + "accuracy": 0.9929719460196793, + "total_bits": 23872512, "q_proj": { "group_size": { "5": 32, @@ -39882,8 +39882,8 @@ } }, { - "accuracy": 0.9884762708097696, - "total_bits": 9727296, + "accuracy": 0.995892624836415, + "total_bits": 29174016, "q_proj": { "group_size": { "6": 128, @@ -39943,8 +39943,8 @@ } }, { - "accuracy": 0.9913079123944044, - "total_bits": 9841920, + "accuracy": 0.9965287849772722, + "total_bits": 29377536, "q_proj": { "group_size": { "6": 32, @@ -40004,8 +40004,8 @@ } }, { - "accuracy": 0.9891001023352146, - "total_bits": 11201856, + "accuracy": 0.9971561338752508, + "total_bits": 33433856, "q_proj": { "group_size": { "6": 128 @@ -40056,8 +40056,8 @@ } }, { - "accuracy": 0.9957126728259027, - "total_bits": 11607296, + "accuracy": 0.9981730832951143, + "total_bits": 34743296, "q_proj": { "group_size": { "6": 32 @@ -40108,8 +40108,8 @@ } }, { - "accuracy": 0.9973605331033468, - "total_bits": 14871872, + "accuracy": 0.9992366975639015, + "total_bits": 44443904, "q_proj": { "group_size": { "8": 128 @@ -40162,8 +40162,8 @@ ], "model.layers.20.mlp": [ { - "accuracy": 0.9159773290157318, - "total_bits": 29656128, + "accuracy": 0.9575619958341122, + "total_bits": 93392016, "gate_proj": { "group_size": { "3": 64, @@ -40214,8 +40214,8 @@ } }, { - "accuracy": 0.9204169064760208, - "total_bits": 30901312, + "accuracy": 0.9590594582259655, + "total_bits": 96259216, "gate_proj": { "group_size": { "3": 64, @@ -40266,8 +40266,8 @@ } }, { - "accuracy": 0.9297328814864159, - "total_bits": 34456000, + "accuracy": 0.9645875468850136, + "total_bits": 107863424, "gate_proj": { "group_size": { "3": 64, @@ -40315,8 +40315,8 @@ } }, { - "accuracy": 0.9330169260501862, - "total_bits": 38584768, + "accuracy": 0.9664024226367474, + "total_bits": 120937856, "gate_proj": { "group_size": { "3": 64, @@ -40364,8 +40364,8 @@ } }, { - "accuracy": 0.9560810998082161, - "total_bits": 43342560, + "accuracy": 0.9781956020742655, + "total_bits": 136162224, "gate_proj": { "group_size": { "4": 128, @@ -40416,8 +40416,8 @@ } }, { - "accuracy": 0.9609145857393742, - "total_bits": 44242688, + "accuracy": 0.9802464488893747, + "total_bits": 139118208, "gate_proj": { "group_size": { "4": 32, @@ -40468,8 +40468,8 @@ } }, { - "accuracy": 0.9647769555449486, - "total_bits": 47520448, + "accuracy": 0.9824922922998667, + "total_bits": 149584176, "gate_proj": { "group_size": { "4": 32, @@ -40517,8 +40517,8 @@ } }, { - "accuracy": 0.9761389158666134, - "total_bits": 54270672, + "accuracy": 0.9883256535977125, + "total_bits": 170943088, "gate_proj": { "group_size": { "4": 128 @@ -40560,8 +40560,8 @@ } }, { - "accuracy": 0.9786328841000795, - "total_bits": 55072512, + "accuracy": 0.989478730596602, + "total_bits": 173467264, "gate_proj": { "group_size": { "4": 32 @@ -40603,8 +40603,8 @@ } }, { - "accuracy": 0.9772225897759199, - "total_bits": 56187616, + "accuracy": 0.9887913670390844, + "total_bits": 176761776, "gate_proj": { "group_size": { "5": 128, @@ -40655,8 +40655,8 @@ } }, { - "accuracy": 0.9807691853493452, - "total_bits": 57087744, + "accuracy": 0.9903437094762921, + "total_bits": 179717760, "gate_proj": { "group_size": { "5": 32, @@ -40707,8 +40707,8 @@ } }, { - "accuracy": 0.9882863629609346, - "total_bits": 69032672, + "accuracy": 0.9942486495710909, + "total_bits": 217361328, "gate_proj": { "group_size": { "6": 128, @@ -40759,8 +40759,8 @@ } }, { - "accuracy": 0.9903248623013496, - "total_bits": 69932800, + "accuracy": 0.9951587240211666, + "total_bits": 220317312, "gate_proj": { "group_size": { "6": 32, @@ -40811,8 +40811,8 @@ } }, { - "accuracy": 0.9931655139662325, - "total_bits": 79551200, + "accuracy": 0.9967739728745073, + "total_bits": 250850224, "gate_proj": { "group_size": { "6": 128 @@ -40854,8 +40854,8 @@ } }, { - "accuracy": 0.9939908175729215, - "total_bits": 82937280, + "accuracy": 0.9970924002118409, + "total_bits": 262906272, "gate_proj": { "group_size": { "8": 128, @@ -40903,8 +40903,8 @@ } }, { - "accuracy": 0.9948418093845248, - "total_bits": 90277312, + "accuracy": 0.9975843168795109, + "total_bits": 286106016, "gate_proj": { "group_size": { "8": 128, @@ -40949,8 +40949,8 @@ } }, { - "accuracy": 0.9981157208094373, - "total_bits": 105219520, + "accuracy": 0.9991227976279333, + "total_bits": 331981216, "gate_proj": { "group_size": { "8": 128 @@ -40991,8 +40991,8 @@ ], "model.layers.21.self_attn": [ { - "accuracy": 0.9216575250029564, - "total_bits": 4051584, + "accuracy": 0.9697561878710985, + "total_bits": 12046848, "q_proj": { "group_size": { "3": 64, @@ -41055,8 +41055,8 @@ } }, { - "accuracy": 0.9397941082715988, - "total_bits": 4199040, + "accuracy": 0.9717134553939104, + "total_bits": 12325376, "q_proj": { "group_size": { "3": 64, @@ -41119,8 +41119,8 @@ } }, { - "accuracy": 0.9431018605828285, - "total_bits": 4293424, + "accuracy": 0.9748040158301592, + "total_bits": 12672960, "q_proj": { "group_size": { "3": 64, @@ -41183,8 +41183,8 @@ } }, { - "accuracy": 0.9477303512394428, - "total_bits": 5070816, + "accuracy": 0.9797326847910881, + "total_bits": 15056256, "q_proj": { "group_size": { "3": 64, @@ -41247,8 +41247,8 @@ } }, { - "accuracy": 0.9575330279767513, - "total_bits": 5958976, + "accuracy": 0.9840058647096157, + "total_bits": 17836288, "q_proj": { "group_size": { "4": 128, @@ -41311,8 +41311,8 @@ } }, { - "accuracy": 0.9581106416881084, - "total_bits": 5962896, + "accuracy": 0.984199883416295, + "total_bits": 17832768, "q_proj": { "group_size": { "4": 128, @@ -41375,8 +41375,8 @@ } }, { - "accuracy": 0.9677457995712757, - "total_bits": 7531840, + "accuracy": 0.9899055035784841, + "total_bits": 22423808, "q_proj": { "group_size": { "4": 128 @@ -41427,8 +41427,8 @@ } }, { - "accuracy": 0.9696231316775084, - "total_bits": 7535760, + "accuracy": 0.9901999542489648, + "total_bits": 22436672, "q_proj": { "group_size": { "4": 128 @@ -41479,8 +41479,8 @@ } }, { - "accuracy": 0.974068894982338, - "total_bits": 7598368, + "accuracy": 0.9908744115382433, + "total_bits": 22623872, "q_proj": { "group_size": { "4": 64 @@ -41531,8 +41531,8 @@ } }, { - "accuracy": 0.9748283289372921, - "total_bits": 7707904, + "accuracy": 0.9917841274291277, + "total_bits": 22946816, "q_proj": { "group_size": { "4": 32 @@ -41583,8 +41583,8 @@ } }, { - "accuracy": 0.9736332800239325, - "total_bits": 7797904, + "accuracy": 0.9920584866777062, + "total_bits": 23337792, "q_proj": { "group_size": { "5": 128, @@ -41647,8 +41647,8 @@ } }, { - "accuracy": 0.9782833456993103, - "total_bits": 7856416, + "accuracy": 0.9924729601480067, + "total_bits": 23303808, "q_proj": { "group_size": { "5": 64, @@ -41711,8 +41711,8 @@ } }, { - "accuracy": 0.9807654041796923, - "total_bits": 7950976, + "accuracy": 0.993480809032917, + "total_bits": 23630336, "q_proj": { "group_size": { "5": 64, @@ -41772,8 +41772,8 @@ } }, { - "accuracy": 0.9832816310226917, - "total_bits": 8006912, + "accuracy": 0.9940678980201483, + "total_bits": 23872512, "q_proj": { "group_size": { "5": 32, @@ -41833,8 +41833,8 @@ } }, { - "accuracy": 0.9884554482996464, - "total_bits": 9727296, + "accuracy": 0.9964499578345567, + "total_bits": 29174016, "q_proj": { "group_size": { "6": 128, @@ -41894,8 +41894,8 @@ } }, { - "accuracy": 0.9900449067354202, - "total_bits": 9841920, + "accuracy": 0.9969881982542574, + "total_bits": 29377536, "q_proj": { "group_size": { "6": 32, @@ -41955,8 +41955,8 @@ } }, { - "accuracy": 0.9905606498941779, - "total_bits": 11201856, + "accuracy": 0.9974862074013799, + "total_bits": 33433856, "q_proj": { "group_size": { "6": 128 @@ -42007,8 +42007,8 @@ } }, { - "accuracy": 0.9957112148404121, - "total_bits": 11607296, + "accuracy": 0.9984284033998847, + "total_bits": 34743296, "q_proj": { "group_size": { "6": 32 @@ -42059,8 +42059,8 @@ } }, { - "accuracy": 0.997682775137946, - "total_bits": 14871872, + "accuracy": 0.9993089917115867, + "total_bits": 44443904, "q_proj": { "group_size": { "8": 128 @@ -42113,8 +42113,8 @@ ], "model.layers.21.mlp": [ { - "accuracy": 0.8147382885217667, - "total_bits": 29656128, + "accuracy": 0.9488564617931843, + "total_bits": 93392016, "gate_proj": { "group_size": { "3": 64, @@ -42165,8 +42165,8 @@ } }, { - "accuracy": 0.8123861104249954, - "total_bits": 30901312, + "accuracy": 0.9508366547524929, + "total_bits": 96259216, "gate_proj": { "group_size": { "3": 64, @@ -42217,8 +42217,8 @@ } }, { - "accuracy": 0.811919167637825, - "total_bits": 34456000, + "accuracy": 0.9570469036698341, + "total_bits": 107863424, "gate_proj": { "group_size": { "3": 64, @@ -42266,8 +42266,8 @@ } }, { - "accuracy": 0.8132558912038803, - "total_bits": 38584768, + "accuracy": 0.9592068418860435, + "total_bits": 120937856, "gate_proj": { "group_size": { "3": 64, @@ -42315,8 +42315,8 @@ } }, { - "accuracy": 0.9520821273326874, - "total_bits": 43342560, + "accuracy": 0.9738200660794973, + "total_bits": 136162224, "gate_proj": { "group_size": { "4": 128, @@ -42367,8 +42367,8 @@ } }, { - "accuracy": 0.9252394661307335, - "total_bits": 44242688, + "accuracy": 0.9763819519430399, + "total_bits": 139118208, "gate_proj": { "group_size": { "4": 32, @@ -42419,8 +42419,8 @@ } }, { - "accuracy": 0.9275848120450974, - "total_bits": 47520448, + "accuracy": 0.97891491279006, + "total_bits": 149584176, "gate_proj": { "group_size": { "4": 32, @@ -42468,8 +42468,8 @@ } }, { - "accuracy": 0.8951225131750107, - "total_bits": 54270672, + "accuracy": 0.9858340658247471, + "total_bits": 170943088, "gate_proj": { "group_size": { "4": 128 @@ -42511,8 +42511,8 @@ } }, { - "accuracy": 0.9324688240885735, - "total_bits": 55072512, + "accuracy": 0.9872882319614291, + "total_bits": 173467264, "gate_proj": { "group_size": { "4": 32 @@ -42554,8 +42554,8 @@ } }, { - "accuracy": 0.9545381739735603, - "total_bits": 56187616, + "accuracy": 0.9865190256386995, + "total_bits": 176761776, "gate_proj": { "group_size": { "5": 128, @@ -42606,8 +42606,8 @@ } }, { - "accuracy": 0.9782946985214949, - "total_bits": 57087744, + "accuracy": 0.9884583679959178, + "total_bits": 179717760, "gate_proj": { "group_size": { "5": 32, @@ -42658,8 +42658,8 @@ } }, { - "accuracy": 0.9647314175963402, - "total_bits": 69032672, + "accuracy": 0.9930870882235467, + "total_bits": 217361328, "gate_proj": { "group_size": { "6": 128, @@ -42710,8 +42710,8 @@ } }, { - "accuracy": 0.9712739903479815, - "total_bits": 69932800, + "accuracy": 0.9942208966240287, + "total_bits": 220317312, "gate_proj": { "group_size": { "6": 32, @@ -42762,8 +42762,8 @@ } }, { - "accuracy": 0.9665376096963882, - "total_bits": 79551200, + "accuracy": 0.9960466707125306, + "total_bits": 250850224, "gate_proj": { "group_size": { "6": 128 @@ -42805,8 +42805,8 @@ } }, { - "accuracy": 0.9931621630676091, - "total_bits": 82937280, + "accuracy": 0.9965014236513525, + "total_bits": 262906272, "gate_proj": { "group_size": { "8": 128, @@ -42854,8 +42854,8 @@ } }, { - "accuracy": 0.9939733603969216, - "total_bits": 90277312, + "accuracy": 0.9970638321246952, + "total_bits": 286106016, "gate_proj": { "group_size": { "8": 128, @@ -42900,8 +42900,8 @@ } }, { - "accuracy": 0.9975375158246607, - "total_bits": 105219520, + "accuracy": 0.9989016323816031, + "total_bits": 331981216, "gate_proj": { "group_size": { "8": 128 @@ -42942,8 +42942,8 @@ ], "model.layers.22.self_attn": [ { - "accuracy": 0.9450578652322292, - "total_bits": 4051584, + "accuracy": 0.960809238255024, + "total_bits": 12046848, "q_proj": { "group_size": { "3": 64, @@ -43006,8 +43006,8 @@ } }, { - "accuracy": 0.9518578052520752, - "total_bits": 4199040, + "accuracy": 0.9627327471971512, + "total_bits": 12325376, "q_proj": { "group_size": { "3": 64, @@ -43070,8 +43070,8 @@ } }, { - "accuracy": 0.9546925686299801, - "total_bits": 4293424, + "accuracy": 0.9660175032913685, + "total_bits": 12672960, "q_proj": { "group_size": { "3": 64, @@ -43134,8 +43134,8 @@ } }, { - "accuracy": 0.9602457955479622, - "total_bits": 5070816, + "accuracy": 0.9719230588525534, + "total_bits": 15056256, "q_proj": { "group_size": { "3": 64, @@ -43198,8 +43198,8 @@ } }, { - "accuracy": 0.9744824357330799, - "total_bits": 5958976, + "accuracy": 0.9805133510380983, + "total_bits": 17836288, "q_proj": { "group_size": { "4": 128, @@ -43262,8 +43262,8 @@ } }, { - "accuracy": 0.9747197534888983, - "total_bits": 5962896, + "accuracy": 0.980793584138155, + "total_bits": 17832768, "q_proj": { "group_size": { "4": 128, @@ -43326,8 +43326,8 @@ } }, { - "accuracy": 0.9786449652165174, - "total_bits": 7531840, + "accuracy": 0.9875983018428087, + "total_bits": 22423808, "q_proj": { "group_size": { "4": 128 @@ -43378,8 +43378,8 @@ } }, { - "accuracy": 0.9789356663823128, - "total_bits": 7535760, + "accuracy": 0.9885378228500485, + "total_bits": 22436672, "q_proj": { "group_size": { "4": 128 @@ -43430,8 +43430,8 @@ } }, { - "accuracy": 0.9807707574218512, - "total_bits": 7598368, + "accuracy": 0.989418632350862, + "total_bits": 22623872, "q_proj": { "group_size": { "4": 64 @@ -43482,8 +43482,8 @@ } }, { - "accuracy": 0.9832558128982782, - "total_bits": 7707904, + "accuracy": 0.9902796112000942, + "total_bits": 22946816, "q_proj": { "group_size": { "4": 32 @@ -43534,8 +43534,8 @@ } }, { - "accuracy": 0.9864223850890994, - "total_bits": 7797904, + "accuracy": 0.9904135735705495, + "total_bits": 23337792, "q_proj": { "group_size": { "5": 128, @@ -43598,8 +43598,8 @@ } }, { - "accuracy": 0.987721985206008, - "total_bits": 7856416, + "accuracy": 0.9907497093081474, + "total_bits": 23303808, "q_proj": { "group_size": { "5": 64, @@ -43662,8 +43662,8 @@ } }, { - "accuracy": 0.9875717340037227, - "total_bits": 7950976, + "accuracy": 0.9913562014698982, + "total_bits": 23630336, "q_proj": { "group_size": { "5": 64, @@ -43723,8 +43723,8 @@ } }, { - "accuracy": 0.9892900278791785, - "total_bits": 8006912, + "accuracy": 0.9928730526007712, + "total_bits": 23872512, "q_proj": { "group_size": { "5": 32, @@ -43784,8 +43784,8 @@ } }, { - "accuracy": 0.9929907601326704, - "total_bits": 9727296, + "accuracy": 0.9956827550195158, + "total_bits": 29174016, "q_proj": { "group_size": { "6": 128, @@ -43845,8 +43845,8 @@ } }, { - "accuracy": 0.9946766439825296, - "total_bits": 9841920, + "accuracy": 0.9962449721060693, + "total_bits": 29377536, "q_proj": { "group_size": { "6": 32, @@ -43906,8 +43906,8 @@ } }, { - "accuracy": 0.9941135626286268, - "total_bits": 11201856, + "accuracy": 0.9967209403403103, + "total_bits": 33433856, "q_proj": { "group_size": { "6": 128 @@ -43958,8 +43958,8 @@ } }, { - "accuracy": 0.9963288505095989, - "total_bits": 11607296, + "accuracy": 0.9980057298671454, + "total_bits": 34743296, "q_proj": { "group_size": { "6": 32 @@ -44010,8 +44010,8 @@ } }, { - "accuracy": 0.9985115990275517, - "total_bits": 14871872, + "accuracy": 0.9991399660357274, + "total_bits": 44443904, "q_proj": { "group_size": { "8": 128 @@ -44064,8 +44064,8 @@ ], "model.layers.22.mlp": [ { - "accuracy": 0.8982994556427002, - "total_bits": 29656128, + "accuracy": 0.9445153661072254, + "total_bits": 93392016, "gate_proj": { "group_size": { "3": 64, @@ -44116,8 +44116,8 @@ } }, { - "accuracy": 0.9027955234050751, - "total_bits": 30901312, + "accuracy": 0.9468347057700157, + "total_bits": 96259216, "gate_proj": { "group_size": { "3": 64, @@ -44168,8 +44168,8 @@ } }, { - "accuracy": 0.9122069776058197, - "total_bits": 34456000, + "accuracy": 0.9540232494473457, + "total_bits": 107863424, "gate_proj": { "group_size": { "3": 64, @@ -44217,8 +44217,8 @@ } }, { - "accuracy": 0.9151279255747795, - "total_bits": 38584768, + "accuracy": 0.9566738121211529, + "total_bits": 120937856, "gate_proj": { "group_size": { "3": 64, @@ -44266,8 +44266,8 @@ } }, { - "accuracy": 0.9454640075564384, - "total_bits": 43342560, + "accuracy": 0.9716828837990761, + "total_bits": 136162224, "gate_proj": { "group_size": { "4": 128, @@ -44318,8 +44318,8 @@ } }, { - "accuracy": 0.9524592384696007, - "total_bits": 44242688, + "accuracy": 0.9743705131113529, + "total_bits": 139118208, "gate_proj": { "group_size": { "4": 32, @@ -44370,8 +44370,8 @@ } }, { - "accuracy": 0.956229392439127, - "total_bits": 47520448, + "accuracy": 0.977362647652626, + "total_bits": 149584176, "gate_proj": { "group_size": { "4": 32, @@ -44419,8 +44419,8 @@ } }, { - "accuracy": 0.9719236381351948, - "total_bits": 54270672, + "accuracy": 0.98453557677567, + "total_bits": 170943088, "gate_proj": { "group_size": { "4": 128 @@ -44462,8 +44462,8 @@ } }, { - "accuracy": 0.9742041788995266, - "total_bits": 55072512, + "accuracy": 0.9861343950033188, + "total_bits": 173467264, "gate_proj": { "group_size": { "4": 32 @@ -44505,8 +44505,8 @@ } }, { - "accuracy": 0.971582468599081, - "total_bits": 56187616, + "accuracy": 0.9854802628979087, + "total_bits": 176761776, "gate_proj": { "group_size": { "5": 128, @@ -44557,8 +44557,8 @@ } }, { - "accuracy": 0.976511787623167, - "total_bits": 57087744, + "accuracy": 0.9875077083706856, + "total_bits": 179717760, "gate_proj": { "group_size": { "5": 32, @@ -44609,8 +44609,8 @@ } }, { - "accuracy": 0.9853576309978962, - "total_bits": 69032672, + "accuracy": 0.9925764654763043, + "total_bits": 217361328, "gate_proj": { "group_size": { "6": 128, @@ -44661,8 +44661,8 @@ } }, { - "accuracy": 0.9882118199020624, - "total_bits": 69932800, + "accuracy": 0.9937437428161502, + "total_bits": 220317312, "gate_proj": { "group_size": { "6": 32, @@ -44713,8 +44713,8 @@ } }, { - "accuracy": 0.9916812535375357, - "total_bits": 79551200, + "accuracy": 0.9957251083105803, + "total_bits": 250850224, "gate_proj": { "group_size": { "6": 128 @@ -44756,8 +44756,8 @@ } }, { - "accuracy": 0.9924999079667032, - "total_bits": 82937280, + "accuracy": 0.9962800405919552, + "total_bits": 262906272, "gate_proj": { "group_size": { "8": 128, @@ -44805,8 +44805,8 @@ } }, { - "accuracy": 0.9932398931123316, - "total_bits": 90277312, + "accuracy": 0.9969717003405094, + "total_bits": 286106016, "gate_proj": { "group_size": { "8": 128, @@ -44851,8 +44851,8 @@ } }, { - "accuracy": 0.9977056230418384, - "total_bits": 105219520, + "accuracy": 0.9988180424552411, + "total_bits": 331981216, "gate_proj": { "group_size": { "8": 128 @@ -44893,8 +44893,8 @@ ], "model.layers.23.self_attn": [ { - "accuracy": 0.9418975599110126, - "total_bits": 4051584, + "accuracy": 0.958304837346077, + "total_bits": 12046848, "q_proj": { "group_size": { "3": 64, @@ -44957,8 +44957,8 @@ } }, { - "accuracy": 0.9507143795490265, - "total_bits": 4199040, + "accuracy": 0.9605084881186485, + "total_bits": 12325376, "q_proj": { "group_size": { "3": 64, @@ -45021,8 +45021,8 @@ } }, { - "accuracy": 0.9567994847893715, - "total_bits": 4293424, + "accuracy": 0.9662404507398605, + "total_bits": 12672960, "q_proj": { "group_size": { "3": 64, @@ -45085,8 +45085,8 @@ } }, { - "accuracy": 0.9656033180654049, - "total_bits": 5070816, + "accuracy": 0.9712386205792427, + "total_bits": 15056256, "q_proj": { "group_size": { "3": 64, @@ -45149,8 +45149,8 @@ } }, { - "accuracy": 0.9721970707178116, - "total_bits": 5958976, + "accuracy": 0.9789633918553591, + "total_bits": 17836288, "q_proj": { "group_size": { "4": 128, @@ -45213,8 +45213,8 @@ } }, { - "accuracy": 0.9728525597602129, - "total_bits": 5962896, + "accuracy": 0.979640532284975, + "total_bits": 17832768, "q_proj": { "group_size": { "4": 128, @@ -45277,8 +45277,8 @@ } }, { - "accuracy": 0.9765131026506424, - "total_bits": 7531840, + "accuracy": 0.983843332156539, + "total_bits": 22423808, "q_proj": { "group_size": { "4": 128 @@ -45329,8 +45329,8 @@ } }, { - "accuracy": 0.9772815871983767, - "total_bits": 7535760, + "accuracy": 0.9846952138468623, + "total_bits": 22436672, "q_proj": { "group_size": { "4": 128 @@ -45381,8 +45381,8 @@ } }, { - "accuracy": 0.9798632990568876, - "total_bits": 7598368, + "accuracy": 0.9862586744129658, + "total_bits": 22623872, "q_proj": { "group_size": { "4": 64 @@ -45433,8 +45433,8 @@ } }, { - "accuracy": 0.9821864292025566, - "total_bits": 7707904, + "accuracy": 0.9882600596174598, + "total_bits": 22946816, "q_proj": { "group_size": { "4": 32 @@ -45485,8 +45485,8 @@ } }, { - "accuracy": 0.9860213585197926, - "total_bits": 7797904, + "accuracy": 0.9891625437885523, + "total_bits": 23337792, "q_proj": { "group_size": { "5": 128, @@ -45549,8 +45549,8 @@ } }, { - "accuracy": 0.987427357584238, - "total_bits": 7856416, + "accuracy": 0.9902145154774189, + "total_bits": 23303808, "q_proj": { "group_size": { "5": 64, @@ -45613,8 +45613,8 @@ } }, { - "accuracy": 0.9871473163366318, - "total_bits": 7950976, + "accuracy": 0.9902783613651991, + "total_bits": 23630336, "q_proj": { "group_size": { "5": 64, @@ -45674,8 +45674,8 @@ } }, { - "accuracy": 0.9887368557974696, - "total_bits": 8006912, + "accuracy": 0.9917801674455404, + "total_bits": 23872512, "q_proj": { "group_size": { "5": 32, @@ -45735,8 +45735,8 @@ } }, { - "accuracy": 0.992696394212544, - "total_bits": 9727296, + "accuracy": 0.9940014062449336, + "total_bits": 29174016, "q_proj": { "group_size": { "6": 128, @@ -45796,8 +45796,8 @@ } }, { - "accuracy": 0.9942985880188644, - "total_bits": 9841920, + "accuracy": 0.9953994629904628, + "total_bits": 29377536, "q_proj": { "group_size": { "6": 32, @@ -45857,8 +45857,8 @@ } }, { - "accuracy": 0.9937034440226853, - "total_bits": 11201856, + "accuracy": 0.9947983603924513, + "total_bits": 33433856, "q_proj": { "group_size": { "6": 128 @@ -45909,8 +45909,8 @@ } }, { - "accuracy": 0.9964339847210795, - "total_bits": 11607296, + "accuracy": 0.9970706596504897, + "total_bits": 34743296, "q_proj": { "group_size": { "6": 32 @@ -45961,8 +45961,8 @@ } }, { - "accuracy": 0.998279616353102, - "total_bits": 14871872, + "accuracy": 0.9986752104014158, + "total_bits": 44443904, "q_proj": { "group_size": { "8": 128 @@ -46015,8 +46015,8 @@ ], "model.layers.23.mlp": [ { - "accuracy": 0.8855618759989738, - "total_bits": 29656128, + "accuracy": 0.9402888938784599, + "total_bits": 93392016, "gate_proj": { "group_size": { "3": 64, @@ -46067,8 +46067,8 @@ } }, { - "accuracy": 0.892217606306076, - "total_bits": 30901312, + "accuracy": 0.94244334846735, + "total_bits": 96259216, "gate_proj": { "group_size": { "3": 64, @@ -46119,8 +46119,8 @@ } }, { - "accuracy": 0.9018212035298347, - "total_bits": 34456000, + "accuracy": 0.9493406303226948, + "total_bits": 107863424, "gate_proj": { "group_size": { "3": 64, @@ -46168,8 +46168,8 @@ } }, { - "accuracy": 0.9049543738365173, - "total_bits": 38584768, + "accuracy": 0.9517794959247112, + "total_bits": 120937856, "gate_proj": { "group_size": { "3": 64, @@ -46217,8 +46217,8 @@ } }, { - "accuracy": 0.9406046271324158, - "total_bits": 43342560, + "accuracy": 0.9693642351776361, + "total_bits": 136162224, "gate_proj": { "group_size": { "4": 128, @@ -46269,8 +46269,8 @@ } }, { - "accuracy": 0.9471482411026955, - "total_bits": 44242688, + "accuracy": 0.9721400607377291, + "total_bits": 139118208, "gate_proj": { "group_size": { "4": 32, @@ -46321,8 +46321,8 @@ } }, { - "accuracy": 0.9510728642344475, - "total_bits": 47520448, + "accuracy": 0.9750117734074593, + "total_bits": 149584176, "gate_proj": { "group_size": { "4": 32, @@ -46370,8 +46370,8 @@ } }, { - "accuracy": 0.9678153768181801, - "total_bits": 54270672, + "accuracy": 0.9836614318192005, + "total_bits": 170943088, "gate_proj": { "group_size": { "4": 128 @@ -46413,8 +46413,8 @@ } }, { - "accuracy": 0.9704958293586969, - "total_bits": 55072512, + "accuracy": 0.9851815141737461, + "total_bits": 173467264, "gate_proj": { "group_size": { "4": 32 @@ -46456,8 +46456,8 @@ } }, { - "accuracy": 0.9693676307797432, - "total_bits": 56187616, + "accuracy": 0.9842824600636959, + "total_bits": 176761776, "gate_proj": { "group_size": { "5": 128, @@ -46508,8 +46508,8 @@ } }, { - "accuracy": 0.9739571679383516, - "total_bits": 57087744, + "accuracy": 0.9864457957446575, + "total_bits": 179717760, "gate_proj": { "group_size": { "5": 32, @@ -46560,8 +46560,8 @@ } }, { - "accuracy": 0.9842521697282791, - "total_bits": 69032672, + "accuracy": 0.9919684743508697, + "total_bits": 217361328, "gate_proj": { "group_size": { "6": 128, @@ -46612,8 +46612,8 @@ } }, { - "accuracy": 0.9868930606171489, - "total_bits": 69932800, + "accuracy": 0.9932071310468018, + "total_bits": 220317312, "gate_proj": { "group_size": { "6": 32, @@ -46664,8 +46664,8 @@ } }, { - "accuracy": 0.9906435487791896, - "total_bits": 79551200, + "accuracy": 0.9954659314826131, + "total_bits": 250850224, "gate_proj": { "group_size": { "6": 128 @@ -46707,8 +46707,8 @@ } }, { - "accuracy": 0.9919623686000705, - "total_bits": 82937280, + "accuracy": 0.9959459374658763, + "total_bits": 262906272, "gate_proj": { "group_size": { "8": 128, @@ -46756,8 +46756,8 @@ } }, { - "accuracy": 0.9928094539791346, - "total_bits": 90277312, + "accuracy": 0.9965702176559716, + "total_bits": 286106016, "gate_proj": { "group_size": { "8": 128, @@ -46802,8 +46802,7812 @@ } }, { - "accuracy": 0.997308891499415, - "total_bits": 105219520, + "accuracy": 0.9987393910996616, + "total_bits": 331981216, + "gate_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + } + ], + "model.layers.24.self_attn": [ + { + "accuracy": 0.9615369476377964, + "total_bits": 12046848, + "q_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9649068638682365, + "total_bits": 12325376, + "q_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9681825265288353, + "total_bits": 12672960, + "q_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9735022038221359, + "total_bits": 15056256, + "q_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9811206813901663, + "total_bits": 17836288, + "q_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9813044182956219, + "total_bits": 17832768, + "q_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "4": 64, + "3": 64 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9856698270887136, + "total_bits": 22423808, + "q_proj": { + "group_size": { + "4": 128 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "4": 128 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "4": 128 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "4": 128 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9859358463436365, + "total_bits": 22436672, + "q_proj": { + "group_size": { + "4": 128 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "4": 128 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "4": 64 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "4": 128 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9884804245084524, + "total_bits": 22623872, + "q_proj": { + "group_size": { + "4": 64 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "4": 64 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "4": 32 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "4": 64 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9889475330710411, + "total_bits": 22946816, + "q_proj": { + "group_size": { + "4": 32 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "4": 32 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "4": 32 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "4": 32 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9901149040088058, + "total_bits": 23337792, + "q_proj": { + "group_size": { + "5": 128, + "4": 128 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "5": 128, + "4": 128 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "5": 64, + "4": 64 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "5": 128, + "4": 128 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9908695789054036, + "total_bits": 23303808, + "q_proj": { + "group_size": { + "5": 64, + "4": 64 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "5": 64, + "4": 64 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "5": 32, + "4": 32 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "5": 64, + "4": 64 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9914190620183945, + "total_bits": 23630336, + "q_proj": { + "group_size": { + "5": 64, + "4": 64 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "5": 64, + "4": 64 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "5": 64 + }, + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "5": 64, + "4": 64 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9922020980156958, + "total_bits": 23872512, + "q_proj": { + "group_size": { + "5": 32, + "4": 32 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "5": 32, + "4": 32 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "5": 32 + }, + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "5": 32, + "4": 32 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9950701901689172, + "total_bits": 29174016, + "q_proj": { + "group_size": { + "6": 128, + "5": 128 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "6": 128, + "5": 128 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "6": 128 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "6": 128, + "5": 128 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9962814920581877, + "total_bits": 29377536, + "q_proj": { + "group_size": { + "6": 32, + "5": 32 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "6": 32, + "5": 32 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "6": 32 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "6": 32, + "5": 32 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9962233561091125, + "total_bits": 33433856, + "q_proj": { + "group_size": { + "6": 128 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "6": 128 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "6": 128 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "6": 128 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9978172562550753, + "total_bits": 34743296, + "q_proj": { + "group_size": { + "6": 32 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "6": 32 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "8": 32 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "6": 32 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9989948330912739, + "total_bits": 44443904, + "q_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + } + ], + "model.layers.24.mlp": [ + { + "accuracy": 0.9374431371688843, + "total_bits": 93392016, + "gate_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "6": 32, + "3": 64, + "2": 64 + }, + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9390751607716084, + "total_bits": 96259216, + "gate_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "6": 32, + "3": 64, + "2": 64 + }, + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9458391033113003, + "total_bits": 107863424, + "gate_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "5": 32, + "3": 32 + }, + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9481438845396042, + "total_bits": 120937856, + "gate_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "5": 32, + "4": 32 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9676165021955967, + "total_bits": 136162224, + "gate_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "4": 128, + "3": 128 + }, + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9704557787626982, + "total_bits": 139118208, + "gate_proj": { + "group_size": { + "4": 32, + "3": 32 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "4": 32, + "3": 32 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "4": 32, + "3": 32 + }, + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9734056629240513, + "total_bits": 149584176, + "gate_proj": { + "group_size": { + "4": 32, + "3": 32 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "4": 32, + "3": 32 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "4": 128 + }, + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9829079657793045, + "total_bits": 170943088, + "gate_proj": { + "group_size": { + "4": 128 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "4": 32 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "4": 128 + }, + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9844172885641456, + "total_bits": 173467264, + "gate_proj": { + "group_size": { + "4": 32 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "4": 32 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "4": 32 + }, + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9833032134920359, + "total_bits": 176761776, + "gate_proj": { + "group_size": { + "5": 128, + "4": 128 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "5": 128, + "4": 128 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "5": 128, + "4": 128 + }, + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9855081383138895, + "total_bits": 179717760, + "gate_proj": { + "group_size": { + "5": 32, + "4": 32 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "5": 32, + "4": 32 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "5": 32, + "4": 32 + }, + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9914482068270445, + "total_bits": 217361328, + "gate_proj": { + "group_size": { + "6": 128, + "5": 128 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "6": 128, + "5": 128 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "6": 128, + "5": 128 + }, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9927391866222024, + "total_bits": 220317312, + "gate_proj": { + "group_size": { + "6": 32, + "5": 32 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "6": 32, + "5": 32 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "6": 32, + "5": 32 + }, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9952755020931363, + "total_bits": 250850224, + "gate_proj": { + "group_size": { + "6": 128 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "6": 128 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "6": 128 + }, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9956771465949714, + "total_bits": 262906272, + "gate_proj": { + "group_size": { + "8": 128, + "6": 128 + }, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "8": 128, + "6": 128 + }, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 128, + "6": 128 + }, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9962611717637628, + "total_bits": 286106016, + "gate_proj": { + "group_size": { + "8": 128, + "6": 128 + }, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "8": 128, + "6": 128 + }, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9987022439017892, + "total_bits": 331981216, + "gate_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + } + ], + "model.layers.25.self_attn": [ + { + "accuracy": 0.9534298852086067, + "total_bits": 12046848, + "q_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9589144624769688, + "total_bits": 12325376, + "q_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9622138701379299, + "total_bits": 12672960, + "q_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9696085806936026, + "total_bits": 15056256, + "q_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9776565134525299, + "total_bits": 17836288, + "q_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9783343803137541, + "total_bits": 17832768, + "q_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "4": 64, + "3": 64 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9821684248745441, + "total_bits": 22423808, + "q_proj": { + "group_size": { + "4": 128 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "4": 128 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "4": 128 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "4": 128 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9827944841235876, + "total_bits": 22436672, + "q_proj": { + "group_size": { + "4": 128 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "4": 128 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "4": 64 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "4": 128 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9850505702197552, + "total_bits": 22623872, + "q_proj": { + "group_size": { + "4": 64 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "4": 64 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "4": 32 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "4": 64 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9865811178460717, + "total_bits": 22946816, + "q_proj": { + "group_size": { + "4": 32 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "4": 32 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "4": 32 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "4": 32 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9889481505379081, + "total_bits": 23337792, + "q_proj": { + "group_size": { + "5": 128, + "4": 128 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "5": 128, + "4": 128 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "5": 64, + "4": 64 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "5": 128, + "4": 128 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9898334480822086, + "total_bits": 23303808, + "q_proj": { + "group_size": { + "5": 64, + "4": 64 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "5": 64, + "4": 64 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "5": 32, + "4": 32 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "5": 64, + "4": 64 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9897657595574856, + "total_bits": 23630336, + "q_proj": { + "group_size": { + "5": 64, + "4": 64 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "5": 64, + "4": 64 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "5": 64 + }, + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "5": 64, + "4": 64 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9907704973593354, + "total_bits": 23872512, + "q_proj": { + "group_size": { + "5": 32, + "4": 32 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "5": 32, + "4": 32 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "5": 32 + }, + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "5": 32, + "4": 32 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9944409807212651, + "total_bits": 29174016, + "q_proj": { + "group_size": { + "6": 128, + "5": 128 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "6": 128, + "5": 128 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "6": 128 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "6": 128, + "5": 128 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9953949535265565, + "total_bits": 29377536, + "q_proj": { + "group_size": { + "6": 32, + "5": 32 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "6": 32, + "5": 32 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "6": 32 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "6": 32, + "5": 32 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9954251786693931, + "total_bits": 33433856, + "q_proj": { + "group_size": { + "6": 128 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "6": 128 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "6": 128 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "6": 128 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9972994164563715, + "total_bits": 34743296, + "q_proj": { + "group_size": { + "6": 32 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "6": 32 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "8": 32 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "6": 32 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.998768886202015, + "total_bits": 44443904, + "q_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + } + ], + "model.layers.25.mlp": [ + { + "accuracy": 0.9338965490460396, + "total_bits": 93392016, + "gate_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "6": 32, + "3": 64, + "2": 64 + }, + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9356790408492088, + "total_bits": 96259216, + "gate_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "6": 32, + "3": 64, + "2": 64 + }, + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9426438584923744, + "total_bits": 107863424, + "gate_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "5": 32, + "3": 32 + }, + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.944975920021534, + "total_bits": 120937856, + "gate_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "5": 32, + "4": 32 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9656462855637074, + "total_bits": 136162224, + "gate_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "4": 128, + "3": 128 + }, + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9688238091766834, + "total_bits": 139118208, + "gate_proj": { + "group_size": { + "4": 32, + "3": 32 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "4": 32, + "3": 32 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "4": 32, + "3": 32 + }, + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9718056470155716, + "total_bits": 149584176, + "gate_proj": { + "group_size": { + "4": 32, + "3": 32 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "4": 32, + "3": 32 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "4": 128 + }, + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9818777497857809, + "total_bits": 170943088, + "gate_proj": { + "group_size": { + "4": 128 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "4": 32 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "4": 128 + }, + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9834900815039873, + "total_bits": 173467264, + "gate_proj": { + "group_size": { + "4": 32 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "4": 32 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "4": 32 + }, + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9823398366570473, + "total_bits": 176761776, + "gate_proj": { + "group_size": { + "5": 128, + "4": 128 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "5": 128, + "4": 128 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "5": 128, + "4": 128 + }, + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9846661388874054, + "total_bits": 179717760, + "gate_proj": { + "group_size": { + "5": 32, + "4": 32 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "5": 32, + "4": 32 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "5": 32, + "4": 32 + }, + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9909223327413201, + "total_bits": 217361328, + "gate_proj": { + "group_size": { + "6": 128, + "5": 128 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "6": 128, + "5": 128 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "6": 128, + "5": 128 + }, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9923099116422236, + "total_bits": 220317312, + "gate_proj": { + "group_size": { + "6": 32, + "5": 32 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "6": 32, + "5": 32 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "6": 32, + "5": 32 + }, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9949582032859325, + "total_bits": 250850224, + "gate_proj": { + "group_size": { + "6": 128 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "6": 128 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "6": 128 + }, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9953958317637444, + "total_bits": 262906272, + "gate_proj": { + "group_size": { + "8": 128, + "6": 128 + }, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "8": 128, + "6": 128 + }, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 128, + "6": 128 + }, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9959803386591375, + "total_bits": 286106016, + "gate_proj": { + "group_size": { + "8": 128, + "6": 128 + }, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "8": 128, + "6": 128 + }, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9986123866401613, + "total_bits": 331981216, + "gate_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + } + ], + "model.layers.26.self_attn": [ + { + "accuracy": 0.9540644697844982, + "total_bits": 12046848, + "q_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.956867054104805, + "total_bits": 12325376, + "q_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9599379301071167, + "total_bits": 12672960, + "q_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9631525836884975, + "total_bits": 15056256, + "q_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9788054525852203, + "total_bits": 17836288, + "q_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9791858848184347, + "total_bits": 17832768, + "q_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "4": 64, + "3": 64 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9820987116545439, + "total_bits": 22423808, + "q_proj": { + "group_size": { + "4": 128 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "4": 128 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "4": 128 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "4": 128 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9823979679495096, + "total_bits": 22436672, + "q_proj": { + "group_size": { + "4": 128 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "4": 128 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "4": 64 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "4": 128 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9851340893656015, + "total_bits": 22623872, + "q_proj": { + "group_size": { + "4": 64 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "4": 64 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "4": 32 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "4": 64 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9859783109277487, + "total_bits": 22946816, + "q_proj": { + "group_size": { + "4": 32 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "4": 32 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "4": 32 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "4": 32 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9895648611709476, + "total_bits": 23337792, + "q_proj": { + "group_size": { + "5": 128, + "4": 128 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "5": 128, + "4": 128 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "5": 64, + "4": 64 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "5": 128, + "4": 128 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9901416758075356, + "total_bits": 23303808, + "q_proj": { + "group_size": { + "5": 64, + "4": 64 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "5": 64, + "4": 64 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "5": 32, + "4": 32 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "5": 64, + "4": 64 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9902024110779166, + "total_bits": 23630336, + "q_proj": { + "group_size": { + "5": 64, + "4": 64 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "5": 64, + "4": 64 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "5": 64 + }, + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "5": 64, + "4": 64 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9905255893245339, + "total_bits": 23872512, + "q_proj": { + "group_size": { + "5": 32, + "4": 32 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "5": 32, + "4": 32 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "5": 32 + }, + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "5": 32, + "4": 32 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9945322521962225, + "total_bits": 29174016, + "q_proj": { + "group_size": { + "6": 128, + "5": 128 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "6": 128, + "5": 128 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "6": 128 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "6": 128, + "5": 128 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9956922074779868, + "total_bits": 29377536, + "q_proj": { + "group_size": { + "6": 32, + "5": 32 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "6": 32, + "5": 32 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "6": 32 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "6": 32, + "5": 32 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9951085280627012, + "total_bits": 33433856, + "q_proj": { + "group_size": { + "6": 128 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "6": 128 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "6": 128 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "6": 128 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9970934621524066, + "total_bits": 34743296, + "q_proj": { + "group_size": { + "6": 32 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "6": 32 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "8": 32 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "6": 32 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9986993519123644, + "total_bits": 44443904, + "q_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + } + ], + "model.layers.26.mlp": [ + { + "accuracy": 0.8959745690226555, + "total_bits": 93392016, + "gate_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "6": 32, + "3": 64, + "2": 64 + }, + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.8906347304582596, + "total_bits": 96259216, + "gate_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "6": 32, + "3": 64, + "2": 64 + }, + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.8982049524784088, + "total_bits": 107863424, + "gate_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "5": 32, + "3": 32 + }, + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9005819037556648, + "total_bits": 120937856, + "gate_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "5": 32, + "4": 32 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9372877180576324, + "total_bits": 136162224, + "gate_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "4": 128, + "3": 128 + }, + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9591436125338078, + "total_bits": 139118208, + "gate_proj": { + "group_size": { + "4": 32, + "3": 32 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "4": 32, + "3": 32 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "4": 32, + "3": 32 + }, + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9634842313826084, + "total_bits": 149584176, + "gate_proj": { + "group_size": { + "4": 32, + "3": 32 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "4": 32, + "3": 32 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "4": 128 + }, + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9737451877444983, + "total_bits": 170943088, + "gate_proj": { + "group_size": { + "4": 128 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "4": 32 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "4": 128 + }, + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9755339920520782, + "total_bits": 173467264, + "gate_proj": { + "group_size": { + "4": 32 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "4": 32 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "4": 32 + }, + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9659285061061382, + "total_bits": 176761776, + "gate_proj": { + "group_size": { + "5": 128, + "4": 128 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "5": 128, + "4": 128 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "5": 128, + "4": 128 + }, + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9786586556583643, + "total_bits": 179717760, + "gate_proj": { + "group_size": { + "5": 32, + "4": 32 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "5": 32, + "4": 32 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "5": 32, + "4": 32 + }, + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9848611913621426, + "total_bits": 217361328, + "gate_proj": { + "group_size": { + "6": 128, + "5": 128 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "6": 128, + "5": 128 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "6": 128, + "5": 128 + }, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.988055014051497, + "total_bits": 220317312, + "gate_proj": { + "group_size": { + "6": 32, + "5": 32 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "6": 32, + "5": 32 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "6": 32, + "5": 32 + }, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9884745078161359, + "total_bits": 250850224, + "gate_proj": { + "group_size": { + "6": 128 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "6": 128 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "6": 128 + }, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9937076321803033, + "total_bits": 262906272, + "gate_proj": { + "group_size": { + "8": 128, + "6": 128 + }, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "8": 128, + "6": 128 + }, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 128, + "6": 128 + }, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9945464618504047, + "total_bits": 286106016, + "gate_proj": { + "group_size": { + "8": 128, + "6": 128 + }, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "8": 128, + "6": 128 + }, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9975181436166167, + "total_bits": 331981216, + "gate_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + } + ], + "model.layers.27.self_attn": [ + { + "accuracy": 0.9458007887005806, + "total_bits": 12046848, + "q_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9537016712129116, + "total_bits": 12325376, + "q_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.95847187936306, + "total_bits": 12672960, + "q_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9664695113897324, + "total_bits": 15056256, + "q_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.974868904799223, + "total_bits": 17836288, + "q_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9752430133521557, + "total_bits": 17832768, + "q_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "4": 64, + "3": 64 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9780112635344267, + "total_bits": 22423808, + "q_proj": { + "group_size": { + "4": 128 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "4": 128 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "4": 128 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "4": 128 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9785104617476463, + "total_bits": 22436672, + "q_proj": { + "group_size": { + "4": 128 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "4": 128 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "4": 64 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "4": 128 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9810489974915981, + "total_bits": 22623872, + "q_proj": { + "group_size": { + "4": 64 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "4": 64 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "4": 32 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "4": 64 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9828283190727234, + "total_bits": 22946816, + "q_proj": { + "group_size": { + "4": 32 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "4": 32 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "4": 32 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "4": 32 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9871984422206879, + "total_bits": 23337792, + "q_proj": { + "group_size": { + "5": 128, + "4": 128 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "5": 128, + "4": 128 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "5": 64, + "4": 64 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "5": 128, + "4": 128 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9880966991186142, + "total_bits": 23303808, + "q_proj": { + "group_size": { + "5": 64, + "4": 64 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "5": 64, + "4": 64 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "5": 32, + "4": 32 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "5": 64, + "4": 64 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9882111232727766, + "total_bits": 23630336, + "q_proj": { + "group_size": { + "5": 64, + "4": 64 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "5": 64, + "4": 64 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "5": 64 + }, + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "5": 64, + "4": 64 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9893290251493454, + "total_bits": 23872512, + "q_proj": { + "group_size": { + "5": 32, + "4": 32 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "5": 32, + "4": 32 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "5": 32 + }, + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "5": 32, + "4": 32 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9933615163899958, + "total_bits": 29174016, + "q_proj": { + "group_size": { + "6": 128, + "5": 128 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "6": 128, + "5": 128 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "6": 128 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "6": 128, + "5": 128 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9945783298462629, + "total_bits": 29377536, + "q_proj": { + "group_size": { + "6": 32, + "5": 32 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "6": 32, + "5": 32 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "6": 32 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "6": 32, + "5": 32 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9940995401702821, + "total_bits": 33433856, + "q_proj": { + "group_size": { + "6": 128 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "6": 128 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "6": 128 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "6": 128 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9963020731229335, + "total_bits": 34743296, + "q_proj": { + "group_size": { + "6": 32 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "6": 32 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "8": 32 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "6": 32 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9984415879007429, + "total_bits": 44443904, + "q_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + } + ], + "model.layers.27.mlp": [ + { + "accuracy": 0.9160272404551506, + "total_bits": 93392016, + "gate_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "6": 32, + "3": 64, + "2": 64 + }, + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9193439707159996, + "total_bits": 96259216, + "gate_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "6": 32, + "3": 64, + "2": 64 + }, + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9267442971467972, + "total_bits": 107863424, + "gate_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "5": 32, + "3": 32 + }, + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9290376827120781, + "total_bits": 120937856, + "gate_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "3": 64, + "2": 64 + }, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "5": 32, + "4": 32 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.955608244985342, + "total_bits": 136162224, + "gate_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "4": 128, + "3": 128 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "4": 128, + "3": 128 + }, + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9611382819712162, + "total_bits": 139118208, + "gate_proj": { + "group_size": { + "4": 32, + "3": 32 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "4": 32, + "3": 32 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "4": 32, + "3": 32 + }, + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9639673866331577, + "total_bits": 149584176, + "gate_proj": { + "group_size": { + "4": 32, + "3": 32 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "4": 32, + "3": 32 + }, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "4": 128 + }, + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9767199419438839, + "total_bits": 170943088, + "gate_proj": { + "group_size": { + "4": 128 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "4": 32 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "4": 128 + }, + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9786195922642946, + "total_bits": 173467264, + "gate_proj": { + "group_size": { + "4": 32 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "4": 32 + }, + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "4": 32 + }, + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9770909082144499, + "total_bits": 176761776, + "gate_proj": { + "group_size": { + "5": 128, + "4": 128 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "5": 128, + "4": 128 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "5": 128, + "4": 128 + }, + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9808260463178158, + "total_bits": 179717760, + "gate_proj": { + "group_size": { + "5": 32, + "4": 32 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "5": 32, + "4": 32 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "5": 32, + "4": 32 + }, + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9881648812443018, + "total_bits": 217361328, + "gate_proj": { + "group_size": { + "6": 128, + "5": 128 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "6": 128, + "5": 128 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "6": 128, + "5": 128 + }, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9903580015525222, + "total_bits": 220317312, + "gate_proj": { + "group_size": { + "6": 32, + "5": 32 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "6": 32, + "5": 32 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "6": 32, + "5": 32 + }, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9931135196238756, + "total_bits": 250850224, + "gate_proj": { + "group_size": { + "6": 128 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "6": 128 + }, + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "6": 128 + }, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9940471947193146, + "total_bits": 262906272, + "gate_proj": { + "group_size": { + "8": 128, + "6": 128 + }, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "8": 128, + "6": 128 + }, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 128, + "6": 128 + }, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9946684688329697, + "total_bits": 286106016, + "gate_proj": { + "group_size": { + "8": 128, + "6": 128 + }, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "8": 128, + "6": 128 + }, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + }, + { + "accuracy": 0.9980629067867994, + "total_bits": 331981216, "gate_proj": { "group_size": { "8": 128 @@ -46847,14 +54651,14 @@ }, "strategy": { "model.layers.0.self_attn": { - "accuracy": 0.9956152755767107, - "total_bits": 11607296, + "accuracy": 0.9954520598985255, + "total_bits": 44443904, "q_proj": { "group_size": { - "6": 32 + "8": 128 }, "bits": [ - 6 + 8 ], "bits_prop": [ 1 @@ -46863,10 +54667,10 @@ }, "k_proj": { "group_size": { - "6": 32 + "8": 128 }, "bits": [ - 6 + 8 ], "bits_prop": [ 1 @@ -46875,7 +54679,7 @@ }, "v_proj": { "group_size": { - "8": 32 + "8": 128 }, "bits": [ 8 @@ -46887,10 +54691,10 @@ }, "o_proj": { "group_size": { - "6": 32 + "8": 128 }, "bits": [ - 6 + 8 ], "bits_prop": [ 1 @@ -46899,8 +54703,8 @@ } }, "model.layers.0.mlp": { - "accuracy": 0.9954390907660127, - "total_bits": 105219520, + "accuracy": 0.997696247883141, + "total_bits": 331981216, "gate_proj": { "group_size": { "8": 128 @@ -46939,14 +54743,14 @@ } }, "model.layers.1.self_attn": { - "accuracy": 0.9939816757105291, - "total_bits": 11201856, + "accuracy": 0.9986925142584369, + "total_bits": 44443904, "q_proj": { "group_size": { - "6": 128 + "8": 128 }, "bits": [ - 6 + 8 ], "bits_prop": [ 1 @@ -46955,10 +54759,10 @@ }, "k_proj": { "group_size": { - "6": 128 + "8": 128 }, "bits": [ - 6 + 8 ], "bits_prop": [ 1 @@ -46967,10 +54771,10 @@ }, "v_proj": { "group_size": { - "6": 128 + "8": 128 }, "bits": [ - 6 + 8 ], "bits_prop": [ 1 @@ -46979,10 +54783,10 @@ }, "o_proj": { "group_size": { - "6": 128 + "8": 128 }, "bits": [ - 6 + 8 ], "bits_prop": [ 1 @@ -46991,8 +54795,8 @@ } }, "model.layers.1.mlp": { - "accuracy": 0.9888631301000714, - "total_bits": 82937280, + "accuracy": 0.9965894345659763, + "total_bits": 286106016, "gate_proj": { "group_size": { "8": 128, @@ -47025,29 +54829,26 @@ }, "down_proj": { "group_size": { - "8": 128, - "6": 128 + "8": 128 }, "bits": [ - 8, - 6 + 8 ], "bits_prop": [ - 0.15, - 0.85 + 1 ], "scale_bits": 4 } }, "model.layers.2.self_attn": { - "accuracy": 0.998056368320249, - "total_bits": 14871872, + "accuracy": 0.997647543437779, + "total_bits": 34743296, "q_proj": { "group_size": { - "8": 128 + "6": 32 }, "bits": [ - 8 + 6 ], "bits_prop": [ 1 @@ -47056,10 +54857,10 @@ }, "k_proj": { "group_size": { - "8": 128 + "6": 32 }, "bits": [ - 8 + 6 ], "bits_prop": [ 1 @@ -47068,7 +54869,7 @@ }, "v_proj": { "group_size": { - "8": 128 + "8": 32 }, "bits": [ 8 @@ -47080,10 +54881,10 @@ }, "o_proj": { "group_size": { - "8": 128 + "6": 32 }, "bits": [ - 8 + 6 ], "bits_prop": [ 1 @@ -47092,8 +54893,8 @@ } }, "model.layers.2.mlp": { - "accuracy": 0.9880842100828886, - "total_bits": 69932800, + "accuracy": 0.9951841593720019, + "total_bits": 220317312, "gate_proj": { "group_size": { "6": 32, @@ -47144,21 +54945,39 @@ } }, "model.layers.3.self_attn": { - "accuracy": 0.9968086224980652, - "total_bits": 11607296, + "accuracy": 0.9981805344577879, + "total_bits": 29377536, "q_proj": { "group_size": { - "6": 32 + "6": 32, + "5": 32 }, "bits": [ - 6 + 6, + 5 ], "bits_prop": [ - 1 + 0.1, + 0.9 ], "scale_bits": 4 }, "k_proj": { + "group_size": { + "6": 32, + "5": 32 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { "group_size": { "6": 32 }, @@ -47170,12 +54989,113 @@ ], "scale_bits": 4 }, + "o_proj": { + "group_size": { + "6": 32, + "5": 32 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + "model.layers.3.mlp": { + "accuracy": 0.9956699670292437, + "total_bits": 217361328, + "gate_proj": { + "group_size": { + "6": 128, + "5": 128 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "6": 128, + "5": 128 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "6": 128, + "5": 128 + }, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "scale_bits": 4 + } + }, + "model.layers.4.self_attn": { + "accuracy": 0.995673674158752, + "total_bits": 23630336, + "q_proj": { + "group_size": { + "5": 64, + "4": 64 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "5": 64, + "4": 64 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, "v_proj": { "group_size": { - "8": 32 + "5": 64 }, "bits": [ - 8 + 5 ], "bits_prop": [ 1 @@ -47184,20 +55104,23 @@ }, "o_proj": { "group_size": { - "6": 32 + "5": 64, + "4": 64 }, "bits": [ - 6 + 5, + 4 ], "bits_prop": [ - 1 + 0.1, + 0.9 ], "scale_bits": 4 } }, - "model.layers.3.mlp": { - "accuracy": 0.9935814472846687, - "total_bits": 69932800, + "model.layers.4.mlp": { + "accuracy": 0.9959681276232004, + "total_bits": 220317312, "gate_proj": { "group_size": { "6": 32, @@ -47247,10 +55170,62 @@ "scale_bits": 4 } }, - "model.layers.4.self_attn": { - "accuracy": 0.9962598541751504, - "total_bits": 9841920, + "model.layers.5.self_attn": { + "accuracy": 0.9995475913165137, + "total_bits": 44443904, "q_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + }, + "model.layers.5.mlp": { + "accuracy": 0.9961935267783701, + "total_bits": 220317312, + "gate_proj": { "group_size": { "6": 32, "5": 32 @@ -47265,11 +55240,63 @@ ], "scale_bits": 4 }, - "k_proj": { + "up_proj": { + "group_size": { + "6": 32, + "5": 32 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { "group_size": { + "8": 32, "6": 32, "5": 32 }, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "scale_bits": 4 + } + }, + "model.layers.6.self_attn": { + "accuracy": 0.9979342515580356, + "total_bits": 29174016, + "q_proj": { + "group_size": { + "6": 128, + "5": 128 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "6": 128, + "5": 128 + }, "bits": [ 6, 5 @@ -47282,7 +55309,7 @@ }, "v_proj": { "group_size": { - "6": 32 + "6": 128 }, "bits": [ 6 @@ -47294,8 +55321,8 @@ }, "o_proj": { "group_size": { - "6": 32, - "5": 32 + "6": 128, + "5": 128 }, "bits": [ 6, @@ -47308,17 +55335,17 @@ "scale_bits": 4 } }, - "model.layers.4.mlp": { - "accuracy": 0.9924559579230845, - "total_bits": 69032672, + "model.layers.6.mlp": { + "accuracy": 0.992395996581763, + "total_bits": 179717760, "gate_proj": { "group_size": { - "6": 128, - "5": 128 + "5": 32, + "4": 32 }, "bits": [ - 6, - 5 + 5, + 4 ], "bits_prop": [ 0.1, @@ -47328,12 +55355,12 @@ }, "up_proj": { "group_size": { - "6": 128, - "5": 128 + "5": 32, + "4": 32 }, "bits": [ - 6, - 5 + 5, + 4 ], "bits_prop": [ 0.25, @@ -47344,13 +55371,13 @@ "down_proj": { "group_size": { "8": 32, - "6": 128, - "5": 128 + "5": 32, + "4": 32 }, "bits": [ 8, - 6, - 5 + 5, + 4 ], "bits_prop": [ 0.05, @@ -47360,9 +55387,9 @@ "scale_bits": 4 } }, - "model.layers.5.self_attn": { - "accuracy": 0.9978703560773283, - "total_bits": 11607296, + "model.layers.7.self_attn": { + "accuracy": 0.9991119984188117, + "total_bits": 34743296, "q_proj": { "group_size": { "6": 32 @@ -47412,13 +55439,13 @@ "scale_bits": 4 } }, - "model.layers.5.mlp": { - "accuracy": 0.9940621736459434, - "total_bits": 69932800, + "model.layers.7.mlp": { + "accuracy": 0.9954051352106035, + "total_bits": 217361328, "gate_proj": { "group_size": { - "6": 32, - "5": 32 + "6": 128, + "5": 128 }, "bits": [ 6, @@ -47432,8 +55459,8 @@ }, "up_proj": { "group_size": { - "6": 32, - "5": 32 + "6": 128, + "5": 128 }, "bits": [ 6, @@ -47448,8 +55475,8 @@ "down_proj": { "group_size": { "8": 32, - "6": 32, - "5": 32 + "6": 128, + "5": 128 }, "bits": [ 8, @@ -47464,15 +55491,15 @@ "scale_bits": 4 } }, - "model.layers.6.self_attn": { - "accuracy": 0.9964089696295559, - "total_bits": 11201856, + "model.layers.8.self_attn": { + "accuracy": 0.9995204001606908, + "total_bits": 44443904, "q_proj": { "group_size": { - "6": 128 + "8": 128 }, "bits": [ - 6 + 8 ], "bits_prop": [ 1 @@ -47481,10 +55508,10 @@ }, "k_proj": { "group_size": { - "6": 128 + "8": 128 }, "bits": [ - 6 + 8 ], "bits_prop": [ 1 @@ -47493,10 +55520,10 @@ }, "v_proj": { "group_size": { - "6": 128 + "8": 128 }, "bits": [ - 6 + 8 ], "bits_prop": [ 1 @@ -47505,7 +55532,105 @@ }, "o_proj": { "group_size": { - "6": 128 + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + }, + "model.layers.8.mlp": { + "accuracy": 0.9927049563266337, + "total_bits": 179717760, + "gate_proj": { + "group_size": { + "5": 32, + "4": 32 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "5": 32, + "4": 32 + }, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 32, + "5": 32, + "4": 32 + }, + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "scale_bits": 4 + } + }, + "model.layers.9.self_attn": { + "accuracy": 0.9980635978281498, + "total_bits": 29377536, + "q_proj": { + "group_size": { + "6": 32, + "5": 32 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "6": 32, + "5": 32 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "6": 32 }, "bits": [ 6 @@ -47514,11 +55639,26 @@ 1 ], "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "6": 32, + "5": 32 + }, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 } }, - "model.layers.6.mlp": { - "accuracy": 0.9879383649677038, - "total_bits": 57087744, + "model.layers.9.mlp": { + "accuracy": 0.9929614635184407, + "total_bits": 179717760, "gate_proj": { "group_size": { "5": 32, @@ -47568,39 +55708,45 @@ "scale_bits": 4 } }, - "model.layers.7.self_attn": { - "accuracy": 0.9971228996291757, - "total_bits": 11607296, + "model.layers.10.self_attn": { + "accuracy": 0.9956478681415319, + "total_bits": 23872512, "q_proj": { "group_size": { - "6": 32 + "5": 32, + "4": 32 }, "bits": [ - 6 + 5, + 4 ], "bits_prop": [ - 1 + 0.1, + 0.9 ], "scale_bits": 4 }, "k_proj": { "group_size": { - "6": 32 + "5": 32, + "4": 32 }, "bits": [ - 6 + 5, + 4 ], "bits_prop": [ - 1 + 0.1, + 0.9 ], "scale_bits": 4 }, "v_proj": { "group_size": { - "8": 32 + "5": 32 }, "bits": [ - 8 + 5 ], "bits_prop": [ 1 @@ -47609,24 +55755,27 @@ }, "o_proj": { "group_size": { - "6": 32 + "5": 32, + "4": 32 }, "bits": [ - 6 + 5, + 4 ], "bits_prop": [ - 1 + 0.1, + 0.9 ], "scale_bits": 4 } }, - "model.layers.7.mlp": { - "accuracy": 0.993465397041291, - "total_bits": 69932800, + "model.layers.10.mlp": { + "accuracy": 0.9959925631992519, + "total_bits": 217361328, "gate_proj": { "group_size": { - "6": 32, - "5": 32 + "6": 128, + "5": 128 }, "bits": [ 6, @@ -47640,8 +55789,8 @@ }, "up_proj": { "group_size": { - "6": 32, - "5": 32 + "6": 128, + "5": 128 }, "bits": [ 6, @@ -47656,8 +55805,8 @@ "down_proj": { "group_size": { "8": 32, - "6": 32, - "5": 32 + "6": 128, + "5": 128 }, "bits": [ 8, @@ -47672,39 +55821,45 @@ "scale_bits": 4 } }, - "model.layers.8.self_attn": { - "accuracy": 0.9987832672195509, - "total_bits": 14871872, + "model.layers.11.self_attn": { + "accuracy": 0.9956227042712271, + "total_bits": 23872512, "q_proj": { "group_size": { - "8": 128 + "5": 32, + "4": 32 }, "bits": [ - 8 + 5, + 4 ], "bits_prop": [ - 1 + 0.1, + 0.9 ], "scale_bits": 4 }, "k_proj": { "group_size": { - "8": 128 + "5": 32, + "4": 32 }, "bits": [ - 8 + 5, + 4 ], "bits_prop": [ - 1 + 0.1, + 0.9 ], "scale_bits": 4 }, "v_proj": { "group_size": { - "8": 128 + "5": 32 }, "bits": [ - 8 + 5 ], "bits_prop": [ 1 @@ -47713,28 +55868,31 @@ }, "o_proj": { "group_size": { - "8": 128 + "5": 32, + "4": 32 }, "bits": [ - 8 + 5, + 4 ], "bits_prop": [ - 1 + 0.1, + 0.9 ], "scale_bits": 4 } }, - "model.layers.8.mlp": { - "accuracy": 0.989203886128962, - "total_bits": 57087744, + "model.layers.11.mlp": { + "accuracy": 0.9968089594040066, + "total_bits": 220317312, "gate_proj": { "group_size": { - "5": 32, - "4": 32 + "6": 32, + "5": 32 }, "bits": [ - 5, - 4 + 6, + 5 ], "bits_prop": [ 0.1, @@ -47744,12 +55902,12 @@ }, "up_proj": { "group_size": { - "5": 32, - "4": 32 + "6": 32, + "5": 32 }, "bits": [ - 5, - 4 + 6, + 5 ], "bits_prop": [ 0.25, @@ -47760,13 +55918,13 @@ "down_proj": { "group_size": { "8": 32, - "5": 32, - "4": 32 + "6": 32, + "5": 32 }, "bits": [ 8, - 5, - 4 + 6, + 5 ], "bits_prop": [ 0.05, @@ -47776,15 +55934,15 @@ "scale_bits": 4 } }, - "model.layers.9.self_attn": { - "accuracy": 0.9963090608362108, - "total_bits": 11607296, + "model.layers.12.self_attn": { + "accuracy": 0.9995390042895451, + "total_bits": 44443904, "q_proj": { "group_size": { - "6": 32 + "8": 128 }, "bits": [ - 6 + 8 ], "bits_prop": [ 1 @@ -47793,10 +55951,10 @@ }, "k_proj": { "group_size": { - "6": 32 + "8": 128 }, "bits": [ - 6 + 8 ], "bits_prop": [ 1 @@ -47805,7 +55963,7 @@ }, "v_proj": { "group_size": { - "8": 32 + "8": 128 }, "bits": [ 8 @@ -47817,10 +55975,10 @@ }, "o_proj": { "group_size": { - "6": 32 + "8": 128 }, "bits": [ - 6 + 8 ], "bits_prop": [ 1 @@ -47828,17 +55986,17 @@ "scale_bits": 4 } }, - "model.layers.9.mlp": { - "accuracy": 0.9933229382149875, - "total_bits": 69032672, + "model.layers.12.mlp": { + "accuracy": 0.9934494704939425, + "total_bits": 179717760, "gate_proj": { "group_size": { - "6": 128, - "5": 128 + "5": 32, + "4": 32 }, "bits": [ - 6, - 5 + 5, + 4 ], "bits_prop": [ 0.1, @@ -47848,12 +56006,12 @@ }, "up_proj": { "group_size": { - "6": 128, - "5": 128 + "5": 32, + "4": 32 }, "bits": [ - 6, - 5 + 5, + 4 ], "bits_prop": [ 0.25, @@ -47864,13 +56022,13 @@ "down_proj": { "group_size": { "8": 32, - "6": 128, - "5": 128 + "5": 32, + "4": 32 }, "bits": [ 8, - 6, - 5 + 5, + 4 ], "bits_prop": [ 0.05, @@ -47880,9 +56038,9 @@ "scale_bits": 4 } }, - "model.layers.10.self_attn": { - "accuracy": 0.9988751251948997, - "total_bits": 14871872, + "model.layers.13.self_attn": { + "accuracy": 0.9995485266845208, + "total_bits": 44443904, "q_proj": { "group_size": { "8": 128 @@ -47932,61 +56090,52 @@ "scale_bits": 4 } }, - "model.layers.10.mlp": { - "accuracy": 0.9905442073941231, - "total_bits": 57087744, + "model.layers.13.mlp": { + "accuracy": 0.9935275404714048, + "total_bits": 173467264, "gate_proj": { "group_size": { - "5": 32, "4": 32 }, "bits": [ - 5, 4 ], "bits_prop": [ - 0.1, - 0.9 + 1 ], "scale_bits": 4 }, "up_proj": { "group_size": { - "5": 32, "4": 32 }, "bits": [ - 5, 4 ], "bits_prop": [ - 0.25, - 0.75 + 1 ], "scale_bits": 4 }, "down_proj": { "group_size": { "8": 32, - "5": 32, "4": 32 }, "bits": [ 8, - 5, 4 ], "bits_prop": [ 0.05, - 0.1, - 0.85 + 0.95 ], "scale_bits": 4 } }, - "model.layers.11.self_attn": { - "accuracy": 0.9981277505867183, - "total_bits": 14871872, + "model.layers.14.self_attn": { + "accuracy": 0.9993740836507641, + "total_bits": 44443904, "q_proj": { "group_size": { "8": 128 @@ -48036,9 +56185,9 @@ "scale_bits": 4 } }, - "model.layers.11.mlp": { - "accuracy": 0.9890436576679349, - "total_bits": 57087744, + "model.layers.14.mlp": { + "accuracy": 0.9935071007348597, + "total_bits": 179717760, "gate_proj": { "group_size": { "5": 32, @@ -48088,9 +56237,9 @@ "scale_bits": 4 } }, - "model.layers.12.self_attn": { - "accuracy": 0.9976966136600822, - "total_bits": 11607296, + "model.layers.15.self_attn": { + "accuracy": 0.9986656652763486, + "total_bits": 34743296, "q_proj": { "group_size": { "6": 32 @@ -48140,91 +56289,88 @@ "scale_bits": 4 } }, - "model.layers.12.mlp": { - "accuracy": 0.990290317684412, - "total_bits": 57087744, + "model.layers.15.mlp": { + "accuracy": 0.992917597759515, + "total_bits": 170943088, "gate_proj": { "group_size": { - "5": 32, - "4": 32 + "4": 128 }, "bits": [ - 5, 4 ], "bits_prop": [ - 0.1, - 0.9 + 1 ], "scale_bits": 4 }, "up_proj": { "group_size": { - "5": 32, "4": 32 }, "bits": [ - 5, 4 ], "bits_prop": [ - 0.25, - 0.75 + 1 ], "scale_bits": 4 }, "down_proj": { "group_size": { "8": 32, - "5": 32, - "4": 32 + "4": 128 }, "bits": [ 8, - 5, 4 ], "bits_prop": [ 0.05, - 0.1, - 0.85 + 0.95 ], "scale_bits": 4 } }, - "model.layers.13.self_attn": { - "accuracy": 0.9970666337758303, - "total_bits": 11607296, + "model.layers.16.self_attn": { + "accuracy": 0.9977092947810888, + "total_bits": 29377536, "q_proj": { "group_size": { - "6": 32 + "6": 32, + "5": 32 }, "bits": [ - 6 + 6, + 5 ], "bits_prop": [ - 1 + 0.1, + 0.9 ], "scale_bits": 4 }, "k_proj": { "group_size": { - "6": 32 + "6": 32, + "5": 32 }, "bits": [ - 6 + 6, + 5 ], "bits_prop": [ - 1 + 0.1, + 0.9 ], "scale_bits": 4 }, "v_proj": { "group_size": { - "8": 32 + "6": 32 }, "bits": [ - 8 + 6 ], "bits_prop": [ 1 @@ -48233,28 +56379,31 @@ }, "o_proj": { "group_size": { - "6": 32 + "6": 32, + "5": 32 }, "bits": [ - 6 + 6, + 5 ], "bits_prop": [ - 1 + 0.1, + 0.9 ], "scale_bits": 4 } }, - "model.layers.13.mlp": { - "accuracy": 0.9945394792594016, - "total_bits": 69932800, + "model.layers.16.mlp": { + "accuracy": 0.9936684546992183, + "total_bits": 179717760, "gate_proj": { "group_size": { - "6": 32, - "5": 32 + "5": 32, + "4": 32 }, "bits": [ - 6, - 5 + 5, + 4 ], "bits_prop": [ 0.1, @@ -48264,12 +56413,12 @@ }, "up_proj": { "group_size": { - "6": 32, - "5": 32 + "5": 32, + "4": 32 }, "bits": [ - 6, - 5 + 5, + 4 ], "bits_prop": [ 0.25, @@ -48280,13 +56429,13 @@ "down_proj": { "group_size": { "8": 32, - "6": 32, - "5": 32 + "5": 32, + "4": 32 }, "bits": [ 8, - 6, - 5 + 5, + 4 ], "bits_prop": [ 0.05, @@ -48296,15 +56445,15 @@ "scale_bits": 4 } }, - "model.layers.14.self_attn": { - "accuracy": 0.9987689683912322, - "total_bits": 14871872, + "model.layers.17.self_attn": { + "accuracy": 0.9988534040749073, + "total_bits": 34743296, "q_proj": { "group_size": { - "8": 128 + "6": 32 }, "bits": [ - 8 + 6 ], "bits_prop": [ 1 @@ -48313,10 +56462,10 @@ }, "k_proj": { "group_size": { - "8": 128 + "6": 32 }, "bits": [ - 8 + 6 ], "bits_prop": [ 1 @@ -48325,7 +56474,7 @@ }, "v_proj": { "group_size": { - "8": 128 + "8": 32 }, "bits": [ 8 @@ -48337,10 +56486,10 @@ }, "o_proj": { "group_size": { - "8": 128 + "6": 32 }, "bits": [ - 8 + 6 ], "bits_prop": [ 1 @@ -48348,9 +56497,9 @@ "scale_bits": 4 } }, - "model.layers.14.mlp": { - "accuracy": 0.9876388935372233, - "total_bits": 55072512, + "model.layers.17.mlp": { + "accuracy": 0.9926247703842819, + "total_bits": 173467264, "gate_proj": { "group_size": { "4": 32 @@ -48391,9 +56540,9 @@ "scale_bits": 4 } }, - "model.layers.15.self_attn": { - "accuracy": 0.9988009865628555, - "total_bits": 14871872, + "model.layers.18.self_attn": { + "accuracy": 0.999480586964637, + "total_bits": 44443904, "q_proj": { "group_size": { "8": 128 @@ -48443,13 +56592,13 @@ "scale_bits": 4 } }, - "model.layers.15.mlp": { - "accuracy": 0.9923870316706598, - "total_bits": 69032672, + "model.layers.18.mlp": { + "accuracy": 0.9962087315507233, + "total_bits": 220317312, "gate_proj": { "group_size": { - "6": 128, - "5": 128 + "6": 32, + "5": 32 }, "bits": [ 6, @@ -48463,8 +56612,8 @@ }, "up_proj": { "group_size": { - "6": 128, - "5": 128 + "6": 32, + "5": 32 }, "bits": [ 6, @@ -48479,8 +56628,8 @@ "down_proj": { "group_size": { "8": 32, - "6": 128, - "5": 128 + "6": 32, + "5": 32 }, "bits": [ 8, @@ -48495,9 +56644,9 @@ "scale_bits": 4 } }, - "model.layers.16.self_attn": { - "accuracy": 0.9953490453772247, - "total_bits": 11607296, + "model.layers.19.self_attn": { + "accuracy": 0.9982008617371321, + "total_bits": 34743296, "q_proj": { "group_size": { "6": 32 @@ -48547,64 +56696,55 @@ "scale_bits": 4 } }, - "model.layers.16.mlp": { - "accuracy": 0.9916866598650813, - "total_bits": 69932800, + "model.layers.19.mlp": { + "accuracy": 0.9964795028790832, + "total_bits": 250850224, "gate_proj": { "group_size": { - "6": 32, - "5": 32 + "6": 128 }, "bits": [ - 6, - 5 + 6 ], "bits_prop": [ - 0.1, - 0.9 + 1 ], "scale_bits": 4 }, "up_proj": { "group_size": { - "6": 32, - "5": 32 + "6": 128 }, "bits": [ - 6, - 5 + 6 ], "bits_prop": [ - 0.25, - 0.75 + 1 ], "scale_bits": 4 }, "down_proj": { "group_size": { "8": 32, - "6": 32, - "5": 32 + "6": 128 }, "bits": [ 8, - 6, - 5 + 6 ], "bits_prop": [ 0.05, - 0.1, - 0.85 + 0.95 ], "scale_bits": 4 } }, - "model.layers.17.self_attn": { - "accuracy": 0.997579267481342, - "total_bits": 11607296, + "model.layers.20.self_attn": { + "accuracy": 0.9971561338752508, + "total_bits": 33433856, "q_proj": { "group_size": { - "6": 32 + "6": 128 }, "bits": [ 6 @@ -48616,7 +56756,7 @@ }, "k_proj": { "group_size": { - "6": 32 + "6": 128 }, "bits": [ 6 @@ -48628,10 +56768,10 @@ }, "v_proj": { "group_size": { - "8": 32 + "6": 128 }, "bits": [ - 8 + 6 ], "bits_prop": [ 1 @@ -48640,7 +56780,7 @@ }, "o_proj": { "group_size": { - "6": 32 + "6": 128 }, "bits": [ 6 @@ -48651,9 +56791,9 @@ "scale_bits": 4 } }, - "model.layers.17.mlp": { - "accuracy": 0.9929212196730077, - "total_bits": 69932800, + "model.layers.20.mlp": { + "accuracy": 0.9951587240211666, + "total_bits": 220317312, "gate_proj": { "group_size": { "6": 32, @@ -48703,9 +56843,9 @@ "scale_bits": 4 } }, - "model.layers.18.self_attn": { - "accuracy": 0.9970566479023546, - "total_bits": 11607296, + "model.layers.21.self_attn": { + "accuracy": 0.9984284033998847, + "total_bits": 34743296, "q_proj": { "group_size": { "6": 32 @@ -48755,61 +56895,52 @@ "scale_bits": 4 } }, - "model.layers.18.mlp": { - "accuracy": 0.9925176077522337, - "total_bits": 69932800, + "model.layers.21.mlp": { + "accuracy": 0.9960466707125306, + "total_bits": 250850224, "gate_proj": { "group_size": { - "6": 32, - "5": 32 + "6": 128 }, "bits": [ - 6, - 5 + 6 ], "bits_prop": [ - 0.1, - 0.9 + 1 ], "scale_bits": 4 }, "up_proj": { "group_size": { - "6": 32, - "5": 32 + "6": 128 }, "bits": [ - 6, - 5 + 6 ], "bits_prop": [ - 0.25, - 0.75 + 1 ], "scale_bits": 4 }, "down_proj": { "group_size": { "8": 32, - "6": 32, - "5": 32 + "6": 128 }, "bits": [ 8, - 6, - 5 + 6 ], "bits_prop": [ 0.05, - 0.1, - 0.85 + 0.95 ], "scale_bits": 4 } }, - "model.layers.19.self_attn": { - "accuracy": 0.9974014775361866, - "total_bits": 11607296, + "model.layers.22.self_attn": { + "accuracy": 0.9980057298671454, + "total_bits": 34743296, "q_proj": { "group_size": { "6": 32 @@ -48859,17 +56990,17 @@ "scale_bits": 4 } }, - "model.layers.19.mlp": { - "accuracy": 0.9905463308095932, - "total_bits": 69932800, + "model.layers.22.mlp": { + "accuracy": 0.9962800405919552, + "total_bits": 262906272, "gate_proj": { "group_size": { - "6": 32, - "5": 32 + "8": 128, + "6": 128 }, "bits": [ - 6, - 5 + 8, + 6 ], "bits_prop": [ 0.1, @@ -48879,47 +57010,44 @@ }, "up_proj": { "group_size": { - "6": 32, - "5": 32 + "8": 128, + "6": 128 }, "bits": [ - 6, - 5 + 8, + 6 ], "bits_prop": [ - 0.25, - 0.75 + 0.1, + 0.9 ], "scale_bits": 4 }, "down_proj": { "group_size": { - "8": 32, - "6": 32, - "5": 32 + "8": 128, + "6": 128 }, "bits": [ 8, - 6, - 5 + 6 ], "bits_prop": [ - 0.05, - 0.1, + 0.15, 0.85 ], "scale_bits": 4 } }, - "model.layers.20.self_attn": { - "accuracy": 0.9957126728259027, - "total_bits": 11607296, + "model.layers.23.self_attn": { + "accuracy": 0.9986752104014158, + "total_bits": 44443904, "q_proj": { "group_size": { - "6": 32 + "8": 128 }, "bits": [ - 6 + 8 ], "bits_prop": [ 1 @@ -48928,10 +57056,10 @@ }, "k_proj": { "group_size": { - "6": 32 + "8": 128 }, "bits": [ - 6 + 8 ], "bits_prop": [ 1 @@ -48940,7 +57068,7 @@ }, "v_proj": { "group_size": { - "8": 32 + "8": 128 }, "bits": [ 8 @@ -48952,10 +57080,10 @@ }, "o_proj": { "group_size": { - "6": 32 + "8": 128 }, "bits": [ - 6 + 8 ], "bits_prop": [ 1 @@ -48963,9 +57091,9 @@ "scale_bits": 4 } }, - "model.layers.20.mlp": { - "accuracy": 0.9903248623013496, - "total_bits": 69932800, + "model.layers.23.mlp": { + "accuracy": 0.9932071310468018, + "total_bits": 220317312, "gate_proj": { "group_size": { "6": 32, @@ -49015,9 +57143,9 @@ "scale_bits": 4 } }, - "model.layers.21.self_attn": { - "accuracy": 0.9957112148404121, - "total_bits": 11607296, + "model.layers.24.self_attn": { + "accuracy": 0.9978172562550753, + "total_bits": 34743296, "q_proj": { "group_size": { "6": 32 @@ -49067,42 +57195,36 @@ "scale_bits": 4 } }, - "model.layers.21.mlp": { - "accuracy": 0.9931621630676091, - "total_bits": 82937280, + "model.layers.24.mlp": { + "accuracy": 0.9952755020931363, + "total_bits": 250850224, "gate_proj": { "group_size": { - "8": 128, "6": 128 }, "bits": [ - 8, 6 ], "bits_prop": [ - 0.1, - 0.9 + 1 ], "scale_bits": 4 }, "up_proj": { "group_size": { - "8": 128, "6": 128 }, "bits": [ - 8, 6 ], "bits_prop": [ - 0.1, - 0.9 + 1 ], "scale_bits": 4 }, "down_proj": { "group_size": { - "8": 128, + "8": 32, "6": 128 }, "bits": [ @@ -49110,21 +57232,21 @@ 6 ], "bits_prop": [ - 0.15, - 0.85 + 0.05, + 0.95 ], "scale_bits": 4 } }, - "model.layers.22.self_attn": { - "accuracy": 0.9985115990275517, - "total_bits": 14871872, + "model.layers.25.self_attn": { + "accuracy": 0.9972994164563715, + "total_bits": 34743296, "q_proj": { "group_size": { - "8": 128 + "6": 32 }, "bits": [ - 8 + 6 ], "bits_prop": [ 1 @@ -49133,10 +57255,10 @@ }, "k_proj": { "group_size": { - "8": 128 + "6": 32 }, "bits": [ - 8 + 6 ], "bits_prop": [ 1 @@ -49145,7 +57267,7 @@ }, "v_proj": { "group_size": { - "8": 128 + "8": 32 }, "bits": [ 8 @@ -49157,10 +57279,10 @@ }, "o_proj": { "group_size": { - "8": 128 + "6": 32 }, "bits": [ - 8 + 6 ], "bits_prop": [ 1 @@ -49168,9 +57290,9 @@ "scale_bits": 4 } }, - "model.layers.22.mlp": { - "accuracy": 0.9882118199020624, - "total_bits": 69932800, + "model.layers.25.mlp": { + "accuracy": 0.9923099116422236, + "total_bits": 220317312, "gate_proj": { "group_size": { "6": 32, @@ -49220,15 +57342,15 @@ "scale_bits": 4 } }, - "model.layers.23.self_attn": { - "accuracy": 0.9964339847210795, - "total_bits": 11607296, + "model.layers.26.self_attn": { + "accuracy": 0.9986993519123644, + "total_bits": 44443904, "q_proj": { "group_size": { - "6": 32 + "8": 128 }, "bits": [ - 6 + 8 ], "bits_prop": [ 1 @@ -49237,10 +57359,10 @@ }, "k_proj": { "group_size": { - "6": 32 + "8": 128 }, "bits": [ - 6 + 8 ], "bits_prop": [ 1 @@ -49249,7 +57371,7 @@ }, "v_proj": { "group_size": { - "8": 32 + "8": 128 }, "bits": [ 8 @@ -49261,20 +57383,118 @@ }, "o_proj": { "group_size": { - "6": 32 + "8": 128 }, "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + }, + "model.layers.26.mlp": { + "accuracy": 0.9945464618504047, + "total_bits": 286106016, + "gate_proj": { + "group_size": { + "8": 128, + "6": 128 + }, + "bits": [ + 8, 6 ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "up_proj": { + "group_size": { + "8": 128, + "6": 128 + }, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + }, + "down_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], "bits_prop": [ 1 ], "scale_bits": 4 } }, - "model.layers.23.mlp": { - "accuracy": 0.9906435487791896, - "total_bits": 79551200, + "model.layers.27.self_attn": { + "accuracy": 0.9984415879007429, + "total_bits": 44443904, + "q_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "k_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "v_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + }, + "o_proj": { + "group_size": { + "8": 128 + }, + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "scale_bits": 4 + } + }, + "model.layers.27.mlp": { + "accuracy": 0.9931135196238756, + "total_bits": 250850224, "gate_proj": { "group_size": { "6": 128 @@ -49316,5 +57536,5 @@ } } }, - "q_last_module_idx": 50 + "q_last_module_idx": 58 } \ No newline at end of file