Upload train_qwen3_codeforces.py with huggingface_hub

train_qwen3_codeforces.py

@@ -17,8 +17,13 @@ Dataset: Competitive programming with chain-of-thought reasoning.
 import trackio
 from datasets import load_dataset
 from peft import LoraConfig
+from transformers import AutoTokenizer
 from trl import SFTTrainer, SFTConfig
 
+# Load tokenizer first to apply chat template
+print("Loading tokenizer...")
+tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-0.6B")
+
 # Load dataset with Python solutions (decontaminated)
 print("Loading dataset open-r1/codeforces-cots...")
 dataset = load_dataset(
@@ -28,6 +33,18 @@ dataset = load_dataset(
 )
 print(f"Dataset loaded: {len(dataset)} examples")
 
+# Formatting function to convert messages to text using chat template
+def formatting_func(example):
+    """Apply chat template to convert messages to text format."""
+    messages = example["messages"]
+    # Apply the tokenizer's chat template
+    text = tokenizer.apply_chat_template(
+        messages,
+        tokenize=False,
+        add_generation_prompt=False
+    )
+    return text
+
 # Create train/eval split
 print("Creating train/eval split...")
 dataset_split = dataset.train_test_split(test_size=0.05, seed=42)
@@ -83,7 +100,7 @@ peft_config = LoraConfig(
     target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
 )
 
-# Initialize trainer
+# Initialize trainer with formatting function
 print("Initializing trainer with Qwen/Qwen3-0.6B...")
 trainer = SFTTrainer(
     model="Qwen/Qwen3-0.6B",
@@ -91,6 +108,7 @@ trainer = SFTTrainer(
     eval_dataset=eval_dataset,
     args=config,
     peft_config=peft_config,
+    formatting_func=formatting_func,
 )
 
 print("Starting training...")