# /// script
# dependencies = ["trl>=0.12.0", "peft>=0.7.0", "datasets>=2.18.0", "trackio"]
# ///

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

model_id = "Qwen/Qwen2.5-0.5B"

# Small subset for a quick test run
dataset = load_dataset("trl-lib/Capybara", split="train")
dataset = dataset.shuffle(seed=42).select(range(500))

# Train/eval split for monitoring
dataset_split = dataset.train_test_split(test_size=0.1, seed=42)

peft_config = LoraConfig(
    r=16,
    lora_alpha=32,
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM",
)

trainer = SFTTrainer(
    model=model_id,
    train_dataset=dataset_split["train"],
    eval_dataset=dataset_split["test"],
    peft_config=peft_config,
    args=SFTConfig(
        output_dir="qwen2.5-0.5b-sft-demo",
        push_to_hub=True,
        hub_model_id="davidsmts/qwen2.5-0.5b-sft-demo",
        hub_strategy="every_save",
        num_train_epochs=1,
        per_device_train_batch_size=1,
        gradient_accumulation_steps=8,
        max_length=512,
        eval_strategy="steps",
        eval_steps=50,
        save_strategy="steps",
        save_steps=100,
        logging_steps=10,
        report_to="trackio",
        project="qwen2.5-sft-demo",
        run_name="qwen2.5-0.5b-capybara",
    ),
)

trainer.train()
trainer.push_to_hub()