backup/delete/eval_mixtral.py

from argparse import ArgumentParser
import numpy as np
import tqdm
import torch
import re
from torch import bfloat16

from transformers import AutoModelForCausalLM, AutoTokenizer
import json
def creat_prompt(model_id, tokenizer=None, question=None, answer=None, pred=None,):
    messages = [
        {
            "role": "system",
            "content":
                "You are an intelligent chatbot designed for evaluating the correctness of generative outputs for question-answer pairs. "
                "Your task is to compare the predicted answer with the correct answer and determine if they match meaningfully. Here's how you can accomplish the task:\n"
                "------\n"
                "##INSTRUCTIONS:\n"
                "- Focus on the meaningful match between the predicted answer and the correct answer.\n"
                "- Consider synonyms or paraphrases as valid matches.\n"
                "- Evaluate the correctness of the prediction compared to the answer."
        },
        {
            "role": "user",
            "content":
                "Please evaluate the following question-answer pair:\n\n"
                f"Question: {question.capitalize()}\n"
                f"Correct Answer: {answer.lower()}\n"
                f"Predicted Answer: {pred.lower()}\n\n"
                "Evaluate if the answer is correct with yes/no and assign a correctness score between 0 and 5, where 0 indicates incorrect answer, and 5 signifies the highest meaningful match. "
                "Please generate the response in the form of a Python dictionary string with keys 'pred' and 'score', where value of 'pred' is  a string of 'yes' or 'no' and value of 'score' is in INTEGER, not STRING. "
                "DO NOT PROVIDE ANY OTHER OUTPUT TEXT OR EXPLANATION. Only provide the Python dictionary string. "
                "For example, your response should look like this: {'pred': 'no', 'score': 0}."
        }
    ]

    if 'mistralai' in model_id:
        prompt = f'<s>[INST] {messages[0]["content"].strip()}\n\n{messages[1]["content"].strip()} [/INST]'
    elif 'NousResearch' in model_id:
        prompt = tokenizer.apply_chat_template(messages, tokenize=False)
        prompt = prompt + '<|im_start|>assistant'
    else:
        raise NotImplementedError
    return prompt

def calculate_ins_level_score(results):
    acc = 0
    ins_num = 0
    for cat_results in results.values():
        acc += cat_results['avg_score'] * cat_results['num_example']
        ins_num += cat_results['num_example']
        return 0
    return acc / ins_num

def LLM_eval(model_id, model, tokenizer, batch_size, samples, cuda=True):

    steps = int(np.ceil(len(samples) / batch_size))
    evals = []
    for step in tqdm.tqdm(range(steps)):
        prompts = []
        for item in samples[step * batch_size: (step + 1) * batch_size]:
            question = item['question'].replace('<image>', '').strip()
            question = question.split('Context:')[0].strip()  # for ScienceQA
            answer = item['answer']
            pred = item['parsed_pred']
            prompt = creat_prompt(model_id, tokenizer, question, answer, pred)
            prompts.append(prompt)

        inputs = tokenizer(prompts, return_tensors="pt", padding=True)  # feed inputs of a batch
        if cuda:
            inputs = {k: v.cuda() for k, v in inputs.items()}
        with torch.inference_mode():
            output_ids = model.generate(
                **inputs,
                use_cache=True,
                max_new_tokens=20,
                pad_token_id=tokenizer.eos_token_id,
            )
        outputs = tokenizer.batch_decode(output_ids[:, inputs['input_ids'].shape[1]:], skip_special_tokens=True) #  decode outputs of a batch
        evals.extend(outputs)
    judge_dict = dict()
    pred_correct = 0
    score_sum = 0

    for sample_data, sample_eval in zip(samples, evals):
        try:
            sample_eval = re.match(r".*(\{.*?\}).*", sample_eval, re.S).group(1)
            sample_eval = sample_eval.replace("'", '"')
            sample_eval = json.loads(sample_eval)
            pred = sample_eval['pred']
            sample_score = sample_eval['score']

            if pred == 'yes':
                judge_dict[sample_data['id']] = {'pred': 'Correct', 'score': sample_score}
                pred_correct += 1
            else:
                judge_dict[sample_data['id']] = {'pred': 'Wrong', 'score': sample_score}
            score_sum += sample_score

        except:
            judge_dict[sample_data['id']] = {'pred': 'Wrong', 'score': 0}

    if len(samples) == 0:
        return {'acc': 0, 'avg_score': 0}
    return judge_dict, {'acc': pred_correct / len(samples), 'avg_score': score_sum / len(samples)}

def run_mixtral_eval():
    tokenizer = AutoTokenizer.from_pretrained(args.model_id)