# /// script
# requires-python = ">=3.10"
# dependencies = [
#     "trl>=0.12.0",
#     "peft>=0.7.0",
#     "transformers>=4.45.0",
#     "datasets>=2.18.0",
#     "accelerate>=0.30.0",
#     "torch>=2.0.0",
# ]
# ///

import os
os.environ["TOKENIZERS_PARALLELISM"] = "false"

from datasets import load_dataset
from peft import LoraConfig
from trl import SFTTrainer, SFTConfig
from transformers import AutoModelForCausalLM, AutoTokenizer

print("Loading tokenizer...")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-0.6B", trust_remote_code=True)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

print("Loading model...")
model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-0.6B",
    torch_dtype="auto",
    trust_remote_code=True,
)

print("Loading dataset...")
dataset = load_dataset("open-r1/codeforces-cots", "solutions_py_decontaminated", split="train")
print(f"Dataset size: {len(dataset)}")

# Take a subset for faster training (full dataset is large)
dataset = dataset.shuffle(seed=42).select(range(min(10000, len(dataset))))
print(f"Using {len(dataset)} examples")

# Split
split = dataset.train_test_split(test_size=0.05, seed=42)
train_dataset = split["train"]
eval_dataset = split["test"]

print("Setting up LoRA...")
peft_config = LoraConfig(
    r=16,
    lora_alpha=32,
    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",
)

print("Setting up training...")
training_args = SFTConfig(
    output_dir="qwen3-0.6b-codeforces-sft",
    push_to_hub=True,
    hub_model_id="luiscosio/qwen3-0.6b-codeforces-sft",
    num_train_epochs=3,
    per_device_train_batch_size=2,
    gradient_accumulation_steps=8,
    gradient_checkpointing=True,
    learning_rate=2e-4,
    lr_scheduler_type="cosine",
    warmup_ratio=0.1,
    eval_strategy="steps",
    eval_steps=100,
    save_strategy="steps",
    save_steps=100,
    save_total_limit=3,
    logging_steps=10,
    bf16=True,
    max_length=2048,
    report_to="none",
)

print("Creating trainer...")
trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    peft_config=peft_config,
    args=training_args,
)

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

print("Saving and pushing to Hub...")
trainer.save_model()
trainer.push_to_hub()

print("Done!")