【Llama 的 o1变体】LLaMA-O1-Supervised-1129 模型：综合指南-CSDN博客

本文链接：https://blog.csdn.net/weixin_41446370/article/details/144322204

简介

LLaMA-O1-Supervised-1129 模型是在 SimpleBerry/OpenLongCoT-SFT 数据集上训练的 LLaMA-O1-Base-1127 模型的微调版本。该模型旨在执行文本生成任务，尤其是在对话环境中。在本文中，我们将探讨该模型的架构、性能，并提供如何有效使用该模型的示例。

在这里插入图片描述

架构

LLaMA-O1-Supervised-1129模型基于LLaMA架构，这是一种基于转换器的语言模型。该模型由编码器和解码器组成，其中编码器接收输入文本并生成连续表示，解码器根据该表示生成输出文本。

LLaMA-O1-Supervised-1129模型已在SimpleBerry/OpenLongCoT-SFT数据集上进行了微调，该数据集是一个包含人类注释文本的大型数据集。微调过程包括调整模型的权重，以优化其在这一特定任务中的性能。

性能

LLaMA-O1-Supervised-1129模型已在文本生成和对话等多项任务中进行了评估。该模型的准确率很高，而且能够生成连贯的、针对特定上下文的文本，显示了良好的效果。

使用示例

要使用 LLaMA-O1-Supervised-1129 模型，您需要安装 Transformers 库。下面是一个如何使用该模型生成文本的示例：

import json
import datasets
import torch
import random
import numpy as np
from transformers import AutoTokenizer, AutoModelForCausalLM



tokenizer = AutoTokenizer.from_pretrained("SimpleBerry/LLaMA-O1-Supervised-1129")
model = AutoModelForCausalLM.from_pretrained("SimpleBerry/LLaMA-O1-Supervised-1129",device_map='auto')



template = "<start_of_father_id>-1<end_of_father_id><start_of_local_id>0<end_of_local_id><start_of_thought><problem>{content}<end_of_thought><start_of_rating><positive_rating><end_of_rating>\n<start_of_father_id>0<end_of_father_id><start_of_local_id>1<end_of_local_id><start_of_thought><expansion>"

def llama_o1_template(query):
    text = template.format(content=query)
    return text


def batch_predict(input_texts):
    input_texts = [input_text.replace('<|end_of_text|>','') for input_text in input_texts]
    inputs = tokenizer(input_texts, return_tensors="pt").to(model.device)
    responses = model.generate(**inputs, max_new_tokens=1024)
    response_texts = tokenizer.batch_decode(responses, skip_special_tokens=False)
    # assitant_responses = [item[len(input_texts[i]):] for i,item in enumerate(response_texts)]
    assitant_responses = [item for i,item in enumerate(response_texts)]
    return assitant_responses

def format_response(response):
    response = response.replace('<start_of_father_id>','')
    response = response.replace('<end_of_father_id><start_of_local_id>','👉')
    response = response.replace('<end_of_local_id><start_of_thought>',', ')
    response = response.replace('<end_of_thought><start_of_rating>','')
    response = response.replace('<end_of_rating>','')
    response = response.replace('<positive_rating>','👍')
    response = response.replace('<negative_rating>','👎')
    return response

i = 'If Diana needs to bike 10 miles to reach home and she can bike at a speed of 3 mph for two hours before getting tired, and then at a speed of 1 mph until she reaches home, how long will it take her to get home?'
input_texts = [llama_o1_template(i)]
assitant_responses = batch_predict(input_texts)
# assitant_responses = [format_response(response) for response in assitant_responses]
print(assitant_responses)