Upload _reasoning_rl_multiseed/eval_reasoning_t2_v7r128_postRL_alpha1_s2_20260507_023054/score.json with huggingface_hub
Browse files
_reasoning_rl_multiseed/eval_reasoning_t2_v7r128_postRL_alpha1_s2_20260507_023054/score.json
ADDED
|
@@ -0,0 +1,71 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"n_rows": 50,
|
| 3 |
+
"format": {
|
| 4 |
+
"reasoning_tags_rate": 0.82,
|
| 5 |
+
"dna_tags_rate": 0.0,
|
| 6 |
+
"cell_mention_rate": 0.9,
|
| 7 |
+
"mean_len_chars": 1022.04,
|
| 8 |
+
"p10_len_chars": 696.1,
|
| 9 |
+
"p90_len_chars": 1554.6
|
| 10 |
+
},
|
| 11 |
+
"tf_grounding": {
|
| 12 |
+
"mean_tf_grounded_rate": 0.33895446884229713,
|
| 13 |
+
"mean_n_cited": 16.22,
|
| 14 |
+
"mean_n_grounded": 5.84,
|
| 15 |
+
"mean_n_canonical_only": 0.16,
|
| 16 |
+
"mean_n_hallucinated": 10.22,
|
| 17 |
+
"mean_motif_consensus": 0.26831356441691256,
|
| 18 |
+
"mean_prog_coverage": 0.1017142857142857
|
| 19 |
+
},
|
| 20 |
+
"per_cell": {
|
| 21 |
+
"Ast": {
|
| 22 |
+
"n": 5,
|
| 23 |
+
"tf_grounded_rate": 0.21537684537684537,
|
| 24 |
+
"n_cited": 11.0,
|
| 25 |
+
"len_chars": 913.6,
|
| 26 |
+
"cell_mention": 0.8
|
| 27 |
+
},
|
| 28 |
+
"End": {
|
| 29 |
+
"n": 3,
|
| 30 |
+
"tf_grounded_rate": 0.19955669955669955,
|
| 31 |
+
"n_cited": 32.0,
|
| 32 |
+
"len_chars": 1380.3333333333333,
|
| 33 |
+
"cell_mention": 1.0
|
| 34 |
+
},
|
| 35 |
+
"Ex": {
|
| 36 |
+
"n": 15,
|
| 37 |
+
"tf_grounded_rate": 0.3518455431037131,
|
| 38 |
+
"n_cited": 18.333333333333332,
|
| 39 |
+
"len_chars": 981.3333333333334,
|
| 40 |
+
"cell_mention": 0.9333333333333333
|
| 41 |
+
},
|
| 42 |
+
"In": {
|
| 43 |
+
"n": 4,
|
| 44 |
+
"tf_grounded_rate": 0.5066823899371069,
|
| 45 |
+
"n_cited": 20.5,
|
| 46 |
+
"len_chars": 914.75,
|
| 47 |
+
"cell_mention": 1.0
|
| 48 |
+
},
|
| 49 |
+
"Mic": {
|
| 50 |
+
"n": 13,
|
| 51 |
+
"tf_grounded_rate": 0.3307791567406952,
|
| 52 |
+
"n_cited": 10.076923076923077,
|
| 53 |
+
"len_chars": 909.6153846153846,
|
| 54 |
+
"cell_mention": 0.8461538461538461
|
| 55 |
+
},
|
| 56 |
+
"OPC": {
|
| 57 |
+
"n": 2,
|
| 58 |
+
"tf_grounded_rate": 0.5875,
|
| 59 |
+
"n_cited": 19.5,
|
| 60 |
+
"len_chars": 1084.0,
|
| 61 |
+
"cell_mention": 1.0
|
| 62 |
+
},
|
| 63 |
+
"Oli": {
|
| 64 |
+
"n": 8,
|
| 65 |
+
"tf_grounded_rate": 0.3115784215784215,
|
| 66 |
+
"n_cited": 16.625,
|
| 67 |
+
"len_chars": 1252.625,
|
| 68 |
+
"cell_mention": 0.875
|
| 69 |
+
}
|
| 70 |
+
}
|
| 71 |
+
}
|