Upload ICM-DPO enhanced Gemma PEFT adapter with comprehensive LoRA and model card

README.md

@@ -50,14 +50,14 @@ This model demonstrates comprehensive capability enhancement using ICM-generated
 ### Dataset
 - **Source**: [codelion/gemma-3-270m-icm-dpo](https://huggingface.co/datasets/codelion/gemma-3-270m-icm-dpo)
 - **Method**: ICM (Internal Coherence Maximization) for label-free preference generation  
-- **Training Samples**: 44286
+- **Training Samples**: 1812
 - **Evaluation Samples**: 50
 
 ### Training Configuration
 - **Epochs**: 1
-- **Batch Size**: 4 (per device)
+- **Batch Size**: 1 (per device)
 - **Gradient Accumulation**: 8 steps
-- **Effective Batch Size**: 32
+- **Effective Batch Size**: 8
 - **Learning Rate**: 2e-07
 - **Optimizer**: paged_adamw_8bit
 - **Memory Optimization**: BF16, Gradient Checkpointing

adapter_config.json

@@ -28,13 +28,13 @@
   "rank_pattern": {},
   "revision": null,
   "target_modules": [
-    "o_proj",
-    "q_proj",
-    "v_proj",
+    "gate_proj",
     "up_proj",
     "k_proj",
+    "v_proj",
     "down_proj",
-    "gate_proj"
+    "o_proj",
+    "q_proj"
   ],
   "target_parameters": null,
   "task_type": "CAUSAL_LM",

adapter_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b937d87b2e5243afade0703add1246907bf08177cd3a0e47f32aefa54da5e594
+oid sha256:8d9fbccbcd31bc8df08fed608b79046818801f5a92ce6537aee6cb7d2b10c085
 size 701497992

added_tokens.json

@@ -0,0 +1,3 @@
+{
+  "<image_soft_token>": 262144
+}

results.json

@@ -1,10 +1,10 @@
 {
   "training_metrics": {
-    "train_runtime": 4688.6863,
-    "train_samples_per_second": 9.445,
-    "train_steps_per_second": 0.295,
+    "train_runtime": 1035.1183,
+    "train_samples_per_second": 1.751,
+    "train_steps_per_second": 0.219,
     "total_flos": 0.0,
-    "train_loss": 0.6801743100833342,
+    "train_loss": 0.6929902463232368,
     "epoch": 1.0
   },
   "config": {
@@ -31,7 +31,7 @@
     "beta": 0.02,
     "max_length": 1024,
     "max_prompt_length": 512,
-    "batch_size": 4,
+    "batch_size": 1,
     "gradient_accumulation_steps": 8,
     "learning_rate": 2e-07,
     "num_train_epochs": 1,

tokenizer.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1299c11d7cf632ef3b4e11937501358ada021bbdf7c47638d13c0ee982f2e79c
+size 4689074

training_config.json

@@ -22,7 +22,7 @@
   "beta": 0.02,
   "max_length": 1024,
   "max_prompt_length": 512,
-  "batch_size": 4,
+  "batch_size": 1,
   "gradient_accumulation_steps": 8,
   "learning_rate": 2e-07,
   "num_train_epochs": 1,