Upload train_qwen_hf_jobs.py with huggingface_hub

train_qwen_hf_jobs.py

@@ -10,6 +10,7 @@ Production-ready script with LoRA, Trackio monitoring, and Hub saving.
 from datasets import load_dataset
 from peft import LoraConfig
 from trl import SFTTrainer, SFTConfig
+from transformers import AutoTokenizer
 import trackio
 
 # Load dataset - using the "messages" field for chat format
@@ -20,6 +21,26 @@ dataset = load_dataset("open-r1/codeforces-cots", "solutions", split="train")
 dataset = dataset.select(range(min(1000, len(dataset))))
 print(f"📊 Training on {len(dataset)} examples")
 
+# Load tokenizer to apply chat template
+tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-0.5B", trust_remote_code=True)
+
+# Define formatting function to convert messages to text using chat template
+def formatting_func(example):
+    """Convert messages field to formatted text using tokenizer's chat template."""
+    if "messages" in example and example["messages"]:
+        # Apply chat template to convert messages to text
+        text = tokenizer.apply_chat_template(
+            example["messages"],
+            tokenize=False,
+            add_generation_prompt=False
+        )
+        return {"text": text}
+    return {"text": ""}
+
+# Apply formatting to dataset
+print("🔄 Formatting dataset with chat template...")
+dataset = dataset.map(formatting_func, remove_columns=dataset.column_names)
+
 # Create train/eval split for monitoring
 dataset_split = dataset.train_test_split(test_size=0.1, seed=42)
 
@@ -41,8 +62,7 @@ trainer = SFTTrainer(
     eval_dataset=dataset_split["test"],
     peft_config=lora_config,
     args=SFTConfig(
-        # Dataset configuration
-        dataset_text_field="messages",  # Specify the messages field for chat format
+        # Dataset configuration - using default "text" field from formatting_func
 
         # Output and Hub settings
         output_dir="qwen-codeforces-sft",