# /// script
# requires-python = ">=3.10"
# dependencies = [
#     "trl>=0.12.0",
#     "peft>=0.7.0",
#     "transformers>=4.45.0",
#     "accelerate>=0.24.0",
#     "trackio",
# ]
# ///

"""
Fine-tune Qwen3-0.6B on open-r1/codeforces-cots for instruction following.
Uses SFT with LoRA for efficient training.
"""

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

# Load dataset - uses messages format which is TRL-compatible
print("Loading open-r1/codeforces-cots dataset...")
dataset = load_dataset("open-r1/codeforces-cots", split="train")
print(f"Dataset loaded: {len(dataset)} examples")

# Sample for faster training (adjust as needed)
if len(dataset) > 10000:
    dataset = dataset.shuffle(seed=42).select(range(10000))
    print(f"Sampled to 10,000 examples for training")

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

# Training configuration
config = SFTConfig(
    # Hub settings
    output_dir="qwen3-0.6b-codeforces-sft",
    push_to_hub=True,
    hub_model_id="atlaswang/qwen3-0.6b-codeforces-sft",
    hub_strategy="every_save",

    # Training parameters
    num_train_epochs=3,
    per_device_train_batch_size=2,
    gradient_accumulation_steps=8,
    learning_rate=2e-4,
    max_length=2048,

    # Logging & checkpointing
    logging_steps=10,
    save_strategy="steps",
    save_steps=200,
    save_total_limit=3,

    # Evaluation
    eval_strategy="steps",
    eval_steps=200,

    # Optimization
    warmup_ratio=0.1,
    lr_scheduler_type="cosine",
    gradient_checkpointing=True,
    bf16=True,

    # Monitoring
    report_to="trackio",
    project="qwen3-codeforces-sft",
    run_name="qwen3-0.6b-codeforces-instruction-tuning",
)

# LoRA configuration
peft_config = LoraConfig(
    r=32,
    lora_alpha=64,
    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"],
)

# Initialize and train
print("Initializing trainer with Qwen/Qwen3-0.6B...")
trainer = SFTTrainer(
    model="Qwen/Qwen3-0.6B",
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    args=config,
    peft_config=peft_config,
)

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

print("Pushing to Hub...")
trainer.push_to_hub()

trackio.finish()

print("Training complete! Model at: https://huggingface.co/atlaswang/qwen3-0.6b-codeforces-sft")