# /// script
# dependencies = [
#     "trl>=0.12.0",
#     "peft>=0.7.0",
#     "transformers>=4.36.0",
#     "accelerate>=0.24.0",
#     "datasets",
#     "torch",
# ]
# ///

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

# Load known-working TRL dataset
print("Loading dataset...")
dataset = load_dataset("trl-lib/Capybara", split="train")
print(f"Dataset loaded: {len(dataset)} examples")

# Small subset for quick test
dataset = dataset.shuffle(seed=42).select(range(1000))
print(f"Using {len(dataset)} examples")

# Split
dataset_split = dataset.train_test_split(test_size=0.1, seed=42)
train_dataset = dataset_split["train"]
eval_dataset = dataset_split["test"]

# Training configuration
config = SFTConfig(
    output_dir="qwen3-0.6b-test",
    push_to_hub=True,
    hub_model_id="luiscosio/qwen3-0.6b-test",
    num_train_epochs=1,
    per_device_train_batch_size=2,
    gradient_accumulation_steps=4,
    gradient_checkpointing=True,
    learning_rate=2e-4,
    logging_steps=10,
    save_strategy="steps",
    save_steps=50,
    eval_strategy="steps",
    eval_steps=50,
    warmup_ratio=0.1,
    bf16=True,
    max_length=1024,
    report_to="none",
)

# 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"],
)

# Initialize and train
print("Initializing trainer...")
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()
print("Done!")