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

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

# Load dataset (ChatML format)
print("Loading pentest dataset...")
dataset = load_dataset(
    "jason-oneal/pentest-agent-dataset",
    data_files="chatml_train.jsonl",
    split="train"
)
print(f"Dataset loaded: {len(dataset)} examples")

# 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"]

# Training configuration
config = SFTConfig(
    output_dir="qwen2.5-coder-1.5b-pentest",
    push_to_hub=True,
    hub_model_id="fawazo/qwen2.5-coder-1.5b-pentest",
    hub_strategy="every_save",
    
    num_train_epochs=3,
    per_device_train_batch_size=4,
    gradient_accumulation_steps=4,
    learning_rate=2e-5,
    
    logging_steps=10,
    save_strategy="steps",
    save_steps=200,
    save_total_limit=2,
    
    eval_strategy="steps",
    eval_steps=200,
    
    warmup_ratio=0.1,
    lr_scheduler_type="cosine",
    
    report_to="trackio",
    project="pentest-coder",
    run_name="qwen2.5-coder-1.5b-sft",
)

# LoRA config for efficient training
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"],
)

# Train
print("Starting training...")
trainer = SFTTrainer(
    model="Qwen/Qwen2.5-Coder-1.5B",
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    args=config,
    peft_config=peft_config,
)

trainer.train()
trainer.push_to_hub()
print("Model saved to: https://huggingface.co/fawazo/qwen2.5-coder-1.5b-pentest")