Upload train_qwen_codeforces.py with huggingface_hub

train_qwen_codeforces.py

@@ -17,12 +17,18 @@ import trackio
 from datasets import load_dataset
 from peft import LoraConfig
 from trl import SFTTrainer, SFTConfig
+from transformers import AutoTokenizer
 
 # Load dataset - using the solutions config with messages column
 print("Loading dataset...")
 dataset = load_dataset("open-r1/codeforces-cots", "solutions", split="train")
 print(f"Dataset loaded: {len(dataset)} examples")
 
+# The dataset has a 'messages' column in chat format
+# We need to keep only the 'messages' column for SFT training
+print("Preparing dataset - keeping only messages column...")
+dataset = dataset.select_columns(["messages"])
+
 # Create train/eval split
 print("Creating train/eval split...")
 dataset_split = dataset.train_test_split(test_size=0.05, seed=42)
@@ -31,6 +37,12 @@ eval_dataset = dataset_split["test"]
 print(f"Train: {len(train_dataset)} examples")
 print(f"Eval: {len(eval_dataset)} examples")
 
+# Load tokenizer for chat template
+print("Loading tokenizer...")
+tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-0.6B")
+if tokenizer.pad_token is None:
+    tokenizer.pad_token = tokenizer.eos_token
+
 # Training configuration
 config = SFTConfig(
     # Hub settings - CRITICAL for saving results
@@ -82,6 +94,7 @@ peft_config = LoraConfig(
 print("Initializing trainer...")
 trainer = SFTTrainer(
     model="Qwen/Qwen3-0.6B",
+    tokenizer=tokenizer,
     train_dataset=train_dataset,
     eval_dataset=eval_dataset,
     args=config,