# /// script
# dependencies = ["trl>=0.20.0", "peft>=0.13.0", "datasets", "transformers>=4.45.0", "accelerate", "bitsandbytes", "huggingface_hub"]
# ///

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

# Authenticate
from huggingface_hub import login
hf_token = os.environ.get("HF_TOKEN")
if hf_token:
    login(token=hf_token)
    print("Authenticated with HuggingFace")

print("Loading dataset...")
dataset = load_dataset("KevinKeller/cognitive-question-generator-v1")
train_dataset = dataset["train"]
eval_dataset = dataset.get("validation")

print(f"Train samples: {len(train_dataset)}")
if eval_dataset:
    print(f"Eval samples: {len(eval_dataset)}")

# Using Qwen2.5-7B for question generation
model_id = "Qwen/Qwen2.5-7B-Instruct"
print(f"Using model: {model_id}")

# LoRA config - slightly higher rank for more complex task
peft_config = LoraConfig(
    r=32,
    lora_alpha=64,
    lora_dropout=0.05,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
    bias="none",
    task_type="CAUSAL_LM",
)

# Training config - modern TRL API
training_args = SFTConfig(
    output_dir="./question-generator-output",
    num_train_epochs=2,
    per_device_train_batch_size=1,
    gradient_accumulation_steps=8,
    learning_rate=1e-4,
    logging_steps=50,
    save_strategy="steps",
    save_steps=500,
    eval_strategy="steps" if eval_dataset else "no",
    eval_steps=500,
    bf16=True,
    push_to_hub=True,
    hub_model_id="KevinKeller/cognitive-question-generator-qwen2.5-7b",
    report_to="none",
    max_length=8192,  # Use max_length, not max_seq_length
)

print("Starting training...")
trainer = SFTTrainer(
    model=model_id,  # Pass model name, not loaded model
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    peft_config=peft_config,
    args=training_args,
)

trainer.train()
print("Training complete! Pushing to Hub...")
trainer.push_to_hub()
print("Done!")