# /// script
# dependencies = [
#     "trl>=0.12.0",
#     "peft>=0.7.0",
#     "transformers>=4.36.0",
#     "datasets>=2.14.0",
# ]
# ///

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

print("🚀 Starting quick demo training...")

# Load a tiny subset of data
dataset = load_dataset("trl-lib/Capybara", split="train[:50]")
print(f"✅ Dataset loaded: {len(dataset)} examples")

# Training configuration
config = SFTConfig(
    # CRITICAL: Hub settings
    output_dir="demo-qwen-sft",
    push_to_hub=True,
    hub_model_id="evalstate/demo-qwen-sft",
    
    # Quick demo settings
    max_steps=10,  # Just 10 steps for quick demo
    per_device_train_batch_size=2,
    learning_rate=2e-5,
    
    # Logging
    logging_steps=2,
    save_strategy="no",  # Don't save checkpoints for quick demo
    
    # Optimization
    warmup_steps=2,
    lr_scheduler_type="constant",
)

# LoRA configuration (reduces memory)
peft_config = LoraConfig(
    r=16,
    lora_alpha=32,
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM",
    target_modules=["q_proj", "v_proj"],
)

# Initialize and train
trainer = SFTTrainer(
    model="Qwen/Qwen2.5-0.5B",
    train_dataset=dataset,
    args=config,
    peft_config=peft_config,
)

print("🏃 Training for 10 steps...")
trainer.train()

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

print("✅ Complete! Model at: https://huggingface.co/evalstate/demo-qwen-sft")