"""SciQ evaluation for quantized MLX LLMs."""

import os
import numpy as np
from tqdm import tqdm
from .common import MultipleChoiceResult, build_multiple_choice_prompt, extract_answer_letter, get_labels

def evaluate_sciq(
    model_path: str,
    n_samples: int = 200,
    seed: int = 42,
) -> MultipleChoiceResult:
    from datasets import load_dataset
    from mlx_lm import load, generate
    from mlx_lm.sample_utils import make_sampler

    if os.path.isdir(model_path):
        model_path = os.path.abspath(model_path)
    model, tokenizer = load(model_path)

    ds = load_dataset("allenai/sciq", split="test")
    
    rng = np.random.RandomState(seed)
    indices = rng.choice(len(ds), size=min(n_samples, len(ds)), replace=False)
    indices.sort()

    n_correct = 0
    per_question = []

    for idx in tqdm(indices, desc="  SciQ eval"):
        item = ds[int(idx)]
        
        question = item["support"] + "\nQuestion: " + item["question"]
        options = [item["distractor1"], item["distractor2"], item["distractor3"], item["correct_answer"]]
        
        state = np.random.RandomState(int(idx))
        order = state.permutation(4)
        shuffled_options = [options[i] for i in order]
        labels = get_labels(4)
        gt_idx = np.where(order == 3)[0][0]
        gt_letter = labels[gt_idx]

        prompt = build_multiple_choice_prompt(question, shuffled_options)

        sampler = make_sampler(temp=0.0)
        output = generate(
            model, tokenizer,
            prompt=prompt,
            max_tokens=15,
            verbose=False,
            sampler=sampler,
        )

        predicted = extract_answer_letter(output, len(shuffled_options))
        correct = predicted == gt_letter
        if correct:
            n_correct += 1

        per_question.append({
            "idx": int(idx),
            "question": item["question"][:100] + "...",
            "ground_truth": gt_letter,
            "predicted": predicted,
            "correct": correct,
        })

    return MultipleChoiceResult(
        n_correct=n_correct,
        n_total=len(indices),
        accuracy=n_correct / len(indices),
        per_question=per_question,
    )