Upload train_qwen3_codeforces.py with huggingface_hub

train_qwen3_codeforces.py

@@ -10,6 +10,12 @@ from trl import SFTTrainer, SFTConfig
 print("Loading dataset...")
 dataset = load_dataset("open-r1/codeforces-cots", "solutions_py_decontaminated", split="train")
 print(f"Dataset size: {len(dataset)} examples")
+print(f"Columns: {dataset.column_names}")
+
+# Check first example to understand structure
+print(f"First example keys: {dataset[0].keys()}")
+if "messages" in dataset.column_names:
+    print(f"Messages sample: {dataset[0]['messages'][:2] if len(dataset[0]['messages']) > 1 else dataset[0]['messages']}")
 
 # Create train/eval split
 dataset_split = dataset.train_test_split(test_size=0.05, seed=42)
@@ -25,7 +31,7 @@ peft_config = LoraConfig(
     task_type="CAUSAL_LM",
 )
 
-# Training config
+# Training config - using "messages" column for chat format
 training_args = SFTConfig(
     output_dir="qwen3-0.6b-codeforces-sft",
     push_to_hub=True,
@@ -50,6 +56,7 @@ training_args = SFTConfig(
     optim="adamw_torch_fused",
     max_grad_norm=1.0,
     max_length=2048,
+    dataset_text_field=None,  # Use messages format
 )
 
 # Initialize trainer