Upload train_qwen_codeforces.py with huggingface_hub

train_qwen_codeforces.py

@@ -1,10 +1,11 @@
 # /// script
-# dependencies = ["trl>=0.12.0", "peft>=0.7.0", "trackio>=0.1.0", "datasets>=2.0.0"]
+# dependencies = ["trl>=0.12.0", "peft>=0.7.0", "trackio>=0.1.0", "datasets>=2.0.0", "transformers>=4.36.0"]
 # ///
 
 from datasets import load_dataset
 from peft import LoraConfig
 from trl import SFTTrainer, SFTConfig
+from transformers import AutoTokenizer
 import trackio
 
 # Load dataset - 1000 examples for ~20 min training
@@ -16,6 +17,23 @@ dataset = load_dataset(
 )
 print(f"📊 Training on {len(dataset)} examples for 3 epochs")
 
+# Load tokenizer to get chat template
+print("🔤 Loading tokenizer...")
+tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-0.5B")
+
+# Define formatting function for messages
+def formatting_func(example):
+    """Convert messages format to text using chat template."""
+    if "messages" in example and example["messages"]:
+        # Use the tokenizer's chat template to format messages
+        text = tokenizer.apply_chat_template(
+            example["messages"],
+            tokenize=False,
+            add_generation_prompt=False
+        )
+        return {"text": text}
+    return {"text": ""}
+
 # LoRA configuration for efficient training
 peft_config = LoraConfig(
     r=8,
@@ -67,6 +85,7 @@ trainer = SFTTrainer(
     train_dataset=dataset,
     args=config,
     peft_config=peft_config,
+    formatting_func=formatting_func,  # Use formatting function for messages
 )
 
 # Train