Upload LoRA model and probe head for run Qwen2.5-7B-Instruct_layer_27

value_head_probes/Qwen2.5-7B-Instruct_layer_27/results.json

@@ -1 +1 @@
-{"eval_metrics": {"eval_accuracy": 0.5802088513177523, "eval_precision": 0.6257100149476831, "eval_recall": 0.4134729355985776, "eval_f1": 0.49791840133222315, "eval_auc": 0.6244539075369286, "eval_lm_loss": 5.963947842337868, "eval_probe_loss": 0.9108255221085115, "eval_sparsity": 0.08431729403409091, "epoch": 4.0}, "train_metrics": {"train_accuracy": 0.9723739129252564, "train_precision": 0.6771163298743856, "train_recall": 0.7932181701855406, "train_f1": 0.7305833824395993, "train_auc": 0.9850187872878208, "train_lm_loss": 5.860074736521795, "train_probe_loss": 0.4101426233513615, "train_sparsity": 0.08815910143729969, "epoch": 4.0}}
+{"eval_metrics": {"eval_accuracy": 0.7532583178175342, "eval_precision": 0.7565023137067177, "eval_recall": 0.7569854702219384, "eval_f1": 0.7567438148443735, "eval_auc": 0.8320366675082659, "eval_lm_loss": 3.6352814110842617, "eval_probe_loss": 0.550237946618687, "eval_sparsity": 0.09692604758522727, "epoch": 2.0}, "train_metrics": {"train_accuracy": 0.9767835414476249, "train_precision": 0.7776664630421503, "train_recall": 0.8277458451455174, "train_f1": 0.8019250636228488, "train_auc": 0.9909954663706676, "train_lm_loss": 3.6980934619903563, "train_probe_loss": 0.42955863563678204, "train_sparsity": 0.08134397849058493, "epoch": 2.0}}

value_head_probes/Qwen2.5-7B-Instruct_layer_27/training_config.json

@@ -7,10 +7,10 @@
     "dtype": "torch.bfloat16",
     "overwrite_output_dir": true,
     "max_steps": -1,
-    "num_train_epochs": 4,
+    "num_train_epochs": 2,
     "per_device_train_batch_size": 2,
     "per_device_eval_batch_size": 2,
-    "lambda_lm": 0.1,
+    "lambda_lm": 0.0,
     "logging_steps": 20,
     "eval_steps": 100,
     "overfit_mode": false,
@@ -42,11 +42,7 @@
         24,
         25,
         26,
-        27,
-        28,
-        29,
-        30,
-        31
+        27
     ],
     "lora_r": 32,
     "lora_alpha": 64,