def gen_sub_dataset():
    # Generate random numbers and their differences
    num_samples = 8000
    rng = np.random.default_rng(42)  # for reproducibility

    # Generate random integers between 0 and 1000
    numbers1 = rng.integers(0, 1000, num_samples)
    numbers2 = rng.integers(0, 1000, num_samples)

    # Create the input strings and differences
    input_pairs = [f"{n2} {n1}" for n1, n2 in zip(numbers1, numbers2)]
    differences = numbers2 - numbers1

    # Create dataset dictionary
    dataset_dict = {
        "input": input_pairs,
        "difference": differences
    }

    # Create and save the dataset
    dataset = Dataset.from_dict(dataset_dict)
    dataset.save_to_disk("subtraction_dataset")

SYSTEM_PROMPT = """ A conversation between User and Assistant. The user gives two numbers, and the Assistant returns the second minus the first number only. The assistant first thinks about the reasoning process in the mind and then provides the user with the answer. The reasoning process and answer are enclosed within <think> </think> and <answer> </answer> tags, respectively, i.e., <think> reasoning process here </think><answer> answer here </answer>
    """

    # Format into conversation
    def make_conversation(example):
        return {
            "prompt": [
                {"role": "system", "content": SYSTEM_PROMPT},
                {"role": "user", "content": example["input"]},
            ]
        }

    dataset = dataset.map(make_conversation)

from transformers import AutoModelForCausalLM
from trl import GRPOConfig, GRPOTrainer
import numpy as np
from datasets import load_from_disk, Dataset
import re
import torch

from peft import LoraConfig

model_id = "Qwen/Qwen2.5-0.5B-Instruct"

import os
os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "1"

model = AutoModelForCausalLM.from_pretrained(
    model_id, 
    device_map="mps", 
    trust_remote_code=True, 
    torch_dtype="auto",  
    low_cpu_mem_usage=True,
    _attn_implementation='eager')

def format_reward(completions, **kwargs):
    """Reward function that checks if the completion has a specific format."""
    pattern = r"^<think>.*?</think><answer>.*?</answer>$"
    completion_contents = [completion[0]["content"] for completion in completions]
    matches = [re.match(pattern, content) for content in completion_contents]
    return [1.0 if match else 0.0 for match in matches]

def accuracy_reward(completions, difference, **kwargs):
    """Reward function that checks if the completion is the same as the ground truth."""
    contents = [completion[0]["content"] for completion in completions]
    answer_pattern = r"<answer>(.*?)</answer>"
    rewards = []
    for content, diff in zip(contents, difference):
        try:
            content = re.search(answer_pattern, content).group(1).strip()
            answer = int(content)
            reward = 1.0 if answer == diff else 0.0
        except Exception:  # if it fails for any reason, return 0.0
            reward = 0.0
        rewards.append(reward)
    return rewards

def finetune():

    training_args = GRPOConfig(
        output_dir="Qwen2-0.5B-GRPO",
        learning_rate=1e-5,
        logging_steps=1,
        gradient_accumulation_steps=8,
        max_completion_length=128,
        per_device_train_batch_size = 1,
        save_steps=10,
        eval_steps=100
    )
    split_dataset = dataset.train_test_split(test_size=0.2, seed=42)
    train_dataset = split_dataset['train']
    test_dataset = split_dataset['test']
     # Initialize the GRPO trainer
    trainer = GRPOTrainer(
        model=model_id,
        reward_funcs=[format_reward, accuracy_reward],
        args=training_args,
        train_dataset=train_dataset,
        eval_dataset=test_dataset,
        peft_config=LoraConfig(task_type="CAUSAL_LM")
    )

    # Train and push the model to the Hub
    trainer.train()

    # Save and push to hub
    trainer.save_model('./output/tunedmodel')