#!/usr/bin/env python3
# /// script
# dependencies = [
#     "trl>=0.12.0",
#     "peft>=0.7.0",
#     "transformers>=4.36.0",
#     "accelerate>=0.24.0",
#     "trackio",
# ]
# ///

"""
SFT training script for Qwen/Qwen2.5-0.5B model.

This script demonstrates:
- Trackio integration for real-time monitoring
- LoRA/PEFT for efficient training
- Proper Hub saving configuration
- Train/eval split for monitoring progress
- Optimized training parameters for small model testing
"""

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

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

# Create train/eval split for monitoring
print("🔀 Creating train/eval split...")
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)} examples")
print(f"   Eval: {len(eval_dataset)} examples")

# Training configuration
print("⚙️  Configuring training parameters...")
config = SFTConfig(
    # CRITICAL: Hub settings - Save model to Hugging Face Hub
    output_dir="qwen-0.5b-sft-capybara",
    push_to_hub=True,
    hub_model_id="vgtomahawk/qwen-0.5b-sft-capybara",
    hub_strategy="every_save",  # Push checkpoints to Hub

    # Training parameters
    num_train_epochs=3,
    per_device_train_batch_size=4,
    gradient_accumulation_steps=4,  # Effective batch size = 4 * 4 = 16
    learning_rate=2e-5,

    # Logging & checkpointing
    logging_steps=10,
    save_strategy="steps",
    save_steps=100,
    save_total_limit=2,  # Keep only last 2 checkpoints

    # Evaluation
    eval_strategy="steps",
    eval_steps=100,

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

    # Monitoring with Trackio
    report_to="trackio",
    project="qwen-sft-demo",
    run_name="qwen-0.5b-baseline",
)

# LoRA configuration for efficient training
print("🔧 Setting up LoRA configuration...")
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 trainer
print("🎯 Initializing SFT trainer...")
trainer = SFTTrainer(
    model="Qwen/Qwen2.5-0.5B",
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    args=config,
    peft_config=peft_config,
)

# Start training
print("🚀 Starting training...")
print("=" * 60)
trainer.train()

# Push final model to Hub
print("=" * 60)
print("💾 Pushing final model to Hub...")
trainer.push_to_hub()

# Complete
print("✅ Training complete!")
print(f"📊 Model available at: https://huggingface.co/vgtomahawk/qwen-0.5b-sft-capybara")
print(f"📈 View training metrics at: https://huggingface.co/spaces/vgtomahawk/trackio")