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

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

# Load the Codeforces CoTs dataset (decontaminated version)
print("Loading dataset...")
dataset = load_dataset("open-r1/codeforces-cots", "solutions_py_decontaminated", split="train")
print(f"Dataset size: {len(dataset)} examples")
print(f"Columns: {dataset.column_names}")

# Check first example to understand structure
print(f"First example keys: {dataset[0].keys()}")
if "messages" in dataset.column_names:
    print(f"Messages sample: {dataset[0]['messages'][:2] if len(dataset[0]['messages']) > 1 else dataset[0]['messages']}")

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

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

# Training config - using "messages" column for chat format
training_args = SFTConfig(
    output_dir="qwen3-0.6b-codeforces-sft",
    push_to_hub=True,
    hub_model_id="luiscosio/qwen3-0.6b-codeforces-sft",
    hub_strategy="every_save",
    num_train_epochs=3,
    per_device_train_batch_size=2,
    gradient_accumulation_steps=8,
    gradient_checkpointing=True,
    learning_rate=2e-4,
    lr_scheduler_type="cosine",
    warmup_ratio=0.1,
    eval_strategy="steps",
    eval_steps=200,
    save_strategy="steps",
    save_steps=200,
    save_total_limit=3,
    logging_steps=10,
    report_to="trackio",
    run_name="qwen3-0.6b-codeforces-sft",
    bf16=True,
    optim="adamw_torch_fused",
    max_grad_norm=1.0,
    max_length=2048,
    dataset_text_field=None,  # Use messages format
)

# Initialize trainer
print("Initializing trainer...")
trainer = SFTTrainer(
    model="Qwen/Qwen3-0.6B",
    train_dataset=dataset_split["train"],
    eval_dataset=dataset_split["test"],
    peft_config=peft_config,
    args=training_args,
)

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

# Push final model to Hub
print("Pushing to Hub...")
trainer.push_to_hub()
print("Training complete!")