train_qwen3_codeforces.py

# /// script
# dependencies = ["trl>=0.12.0", "peft>=0.7.0", "trackio", "transformers>=4.44.0", "datasets>=2.14.0", "torch>=2.0.0", "accelerate>=0.24.0"]
# ///

"""
Fine-tune Qwen3-0.6B on open-r1/codeforces-cots for instruction following.
This script uses SFT (Supervised Fine-Tuning) with LoRA for efficient training.
"""

from datasets import load_dataset
from peft import LoraConfig
from trl import SFTTrainer, SFTConfig
from transformers import AutoTokenizer
import trackio

print("Loading model and tokenizer...")
model_name = "Qwen/Qwen2.5-0.5B"  # Using Qwen2.5-0.5B as base (closest to 0.6B)
tokenizer = AutoTokenizer.from_pretrained(model_name)

print("Loading dataset...")
dataset = load_dataset("open-r1/codeforces-cots", name="solutions_py_decontaminated", split="train")

print(f"Dataset size: {len(dataset)}")

# Take a manageable subset for initial training
dataset = dataset.select(range(min(5000, len(dataset))))
print(f"Using {len(dataset)} examples for training")

# Filter out examples without valid messages
print("Filtering valid examples...")
dataset = dataset.filter(lambda x: x.get('messages') and len(x['messages']) >= 2)
print(f"After filtering: {len(dataset)} examples")

# Create train/eval split for monitoring
dataset_split = dataset.train_test_split(test_size=0.05, seed=42)
print(f"Train: {len(dataset_split['train'])}, Eval: {len(dataset_split['test'])}")

# Configure LoRA for efficient fine-tuning
peft_config = LoraConfig(
    r=16,
    lora_alpha=32,
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM",
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]
)

# Training configuration
training_args = SFTConfig(
    output_dir="qwen3-codeforces-sft",

    # Training parameters
    num_train_epochs=3,
    per_device_train_batch_size=2,
    per_device_eval_batch_size=2,
    gradient_accumulation_steps=8,  # Effective batch size = 16
    gradient_checkpointing=True,

    # Optimization
    learning_rate=2e-4,
    lr_scheduler_type="cosine",
    warmup_ratio=0.1,
    optim="adamw_torch",

    # Evaluation and saving
    eval_strategy="steps",
    eval_steps=100,
    save_strategy="steps",
    save_steps=200,
    save_total_limit=3,

    # Logging
    logging_steps=10,
    report_to="trackio",

    # Hub configuration - CRITICAL for saving results
    push_to_hub=True,
    hub_model_id="udaykiran212/qwen3-0.6b-codeforces-sft",
    hub_strategy="every_save",
    hub_private_repo=False,

    # Performance
    bf16=True,
    max_grad_norm=1.0,

    # Dataset formatting - NO dataset_text_field for chat format
    packing=False,
)

# Initialize trainer for chat format
print("Initializing trainer...")
trainer = SFTTrainer(
    model=model_name,
    train_dataset=dataset_split["train"],
    eval_dataset=dataset_split["test"],
    peft_config=peft_config,
    args=training_args,
    processing_class=tokenizer,
)

# Start training
print("Starting training...")
trainer.train()

# Save and push final model
print("Saving final model...")
trainer.save_model()
trainer.push_to_hub()

print("Training completed successfully!")
print(f"Model saved to: https://huggingface.co/udaykiran212/qwen3-0.6b-codeforces-sft")