# /// script
# dependencies = [
#     "trl>=0.12.0",
#     "peft>=0.7.0",
#     "transformers>=4.51.0",
#     "accelerate>=0.24.0",
#     "trackio",
#     "datasets",
# ]
# ///

"""
SFT training: Qwen3-0.6B on open-r1/codeforces-cots for instruction following.
Quick test run with 100 examples.
"""

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

# Load dataset - using solutions_w_editorials config for instruction following
print("Loading dataset...")
dataset = load_dataset(
    "open-r1/codeforces-cots",
    "solutions_w_editorials",
    split="train"
)
print(f"Full dataset: {len(dataset)} examples")

# Take 100 examples for quick test
dataset = dataset.select(range(min(100, len(dataset))))
print(f"Using {len(dataset)} examples for training")

# Create train/eval split (90/10)
dataset_split = dataset.train_test_split(test_size=0.1, seed=42)
train_dataset = dataset_split["train"]
eval_dataset = dataset_split["test"]
print(f"Train: {len(train_dataset)}, Eval: {len(eval_dataset)}")

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

    # Training parameters
    num_train_epochs=1,
    per_device_train_batch_size=2,
    gradient_accumulation_steps=4,
    learning_rate=2e-4,
    max_length=2048,

    # Logging & checkpoints
    logging_steps=5,
    save_strategy="steps",
    save_steps=50,
    save_total_limit=2,

    # Evaluation
    eval_strategy="steps",
    eval_steps=25,

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

    # Monitoring
    report_to="trackio",
    project="qwen3-codeforces",
    run_name="sft-test-100examples",
)

# LoRA configuration
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"],
)

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

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

# Push to Hub
print("Pushing model to Hub...")
trainer.push_to_hub()

print("Training complete!")
print(f"Model: https://huggingface.co/gilbaes/qwen3-0.6b-codeforces-sft")
print(f"Trackio: https://huggingface.co/spaces/gilbaes/trackio")