Update with a retrain at ~1000 steps with fixed QLoRA repo.

adapter_config.json

@@ -4,19 +4,19 @@
   "fan_in_fan_out": false,
   "inference_mode": true,
   "init_lora_weights": true,
-  "lora_alpha": 16,
-  "lora_dropout": 0.0,
+  "lora_alpha": 16.0,
+  "lora_dropout": 0.05,
   "modules_to_save": null,
   "peft_type": "LORA",
   "r": 64,
   "target_modules": [
-    "up_proj",
-    "k_proj",
-    "v_proj",
     "gate_proj",
-    "o_proj",
     "q_proj",
-    "down_proj"
+    "down_proj",
+    "v_proj",
+    "o_proj",
+    "k_proj",
+    "up_proj"
   ],
   "task_type": "CAUSAL_LM"
 }

adapter_model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6bb4ae7b8f1cad64867c2baaebeda6f25f573cb8f04598899f08f8e902b18fc4
-size 3198561797
+oid sha256:8d654274c87e6a31e2db91ca80ac2860259657f680821526cfbb58697e893990
+size 1599483397