Refresh base adapter to best eval_loss (profile=llama3_2_3b_base, step=216, eval_loss=0.052900)

meta-llama__llama-3.2-3b/adapter_config.json

@@ -29,17 +29,17 @@
   "peft_type": "LORA",
   "peft_version": "0.18.1",
   "qalora_group_size": 16,
-  "r": 8,
+  "r": 16,
   "rank_pattern": {},
   "revision": null,
   "target_modules": [
-    "q_proj",
+    "up_proj",
     "down_proj",
-    "k_proj",
-    "o_proj",
     "v_proj",
+    "q_proj",
     "gate_proj",
-    "up_proj"
+    "o_proj",
+    "k_proj"
   ],
   "target_parameters": null,
   "task_type": "CAUSAL_LM",

meta-llama__llama-3.2-3b/adapter_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5da5f577d1bf82f167b478346812d5d47b4e7584dc622e4048cc3288a7fcdb19
-size 48679352
+oid sha256:3af92d0db3c1ac3a17f1e1e764d362904fd7d8bb821668eb2fc4f4154dc2926a
+size 97307544

meta-llama__llama-3.2-3b/best_eval_metrics.json

@@ -1,5 +1,5 @@
 {
-  "best_eval_loss": 0.048659659922122955,
-  "best_step": 1152,
-  "epoch": 2.0034782608695654
+  "best_eval_loss": 0.05289991945028305,
+  "best_step": 216,
+  "epoch": 0.7526132404181185
 }