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

"""
Fine-tune Qwen3-8B on Vyvo Life CoPilot conversations dataset.
"""

import json
from datasets import Dataset
from peft import LoraConfig
from trl import SFTTrainer, SFTConfig
from huggingface_hub import hf_hub_download

print("📦 Downloading dataset from Hub...")
data_path = hf_hub_download(
    repo_id="Codyfederer/vyvo-text-conversations",
    filename="text_conversations.jsonl",
    repo_type="dataset"
)

# Load JSONL manually to avoid schema inference issues
print("🔄 Loading and converting to messages format...")
conversations = []
with open(data_path, 'r', encoding='utf-8') as f:
    for line in f:
        item = json.loads(line)
        messages = []
        for turn in item['turns']:
            messages.append({
                'role': turn['role'],
                'content': turn['content']
            })
        conversations.append({'messages': messages})

dataset = Dataset.from_list(conversations)
print(f"✅ Converted {len(dataset)} conversations")

# Create train/eval split
print("🔀 Creating train/eval split...")
dataset_split = dataset.train_test_split(test_size=0.05, 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 - optimized for memory on A10G
config = SFTConfig(
    # Hub settings
    output_dir="qwen3-8b-vyvo-copilot",
    push_to_hub=True,
    hub_model_id="Codyfederer/qwen3-8b-vyvo-copilot",
    hub_strategy="every_save",
    hub_private_repo=False,

    # Training parameters - reduced for memory
    num_train_epochs=3,
    per_device_train_batch_size=1,  # Reduced from 2
    gradient_accumulation_steps=16,  # Increased to maintain effective batch size
    learning_rate=2e-4,
    max_length=1024,  # Reduced from 2048 to save memory

    # Memory optimization
    gradient_checkpointing=True,
    gradient_checkpointing_kwargs={"use_reentrant": False},
    bf16=True,
    optim="adamw_8bit",  # Use 8-bit optimizer to save memory

    # Logging & checkpointing
    logging_steps=10,
    save_strategy="steps",
    save_steps=200,
    save_total_limit=2,

    # Evaluation - skip eval during training to save memory
    eval_strategy="no",

    # Optimization
    warmup_ratio=0.05,
    lr_scheduler_type="cosine",
    weight_decay=0.01,

    # Monitoring
    report_to="trackio",
    project="vyvo-copilot-training",
    run_name="qwen3-8b-sft-v1",
)

# LoRA configuration - optimized for Qwen3
peft_config = LoraConfig(
    r=32,
    lora_alpha=64,
    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 and train
print("🎯 Initializing trainer with Qwen/Qwen3-8B...")
trainer = SFTTrainer(
    model="Qwen/Qwen3-8B",
    train_dataset=train_dataset,
    args=config,
    peft_config=peft_config,
)

print("🚀 Starting training...")
trainer.train()

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

print("✅ Training complete!")
print("📦 Model saved to: https://huggingface.co/Codyfederer/qwen3-8b-vyvo-copilot")
print("📊 View metrics at: https://huggingface.co/spaces/Codyfederer/trackio")