大模型微调示例四之Llama-Factory-DPO

一、强化学习数据处理

原始数据地址：https://nijianmo.github.io/amazon/index.html

第一步：读取 video game 信息

import codecs, json, re
from random import shuffle# 第一步：读取 video game 信息
# key 是 productID，value是 title
games = {}
cc = 0with codecs.open('./data/src_data/meta_Video_Games.json', mode='r') as fin:for line in fin:tmp_info = json.loads(line.strip())# asin - ID of the product# title - name of the productgames[tmp_info["asin"]] = tmp_info["title"]if len(games) % 10000 == 0:print(f'Length of games: {len(games)}')

第二步：读取用户评分信息

# key 是 userid，value 是评价的游戏和评分
user_reviews = {}cc = 0
with codecs.open('./data/src_data/Video_Games_5.json', mode='r') as fin:for line in fin:tmp_info = json.loads(line.strip())# reviewerID - ID of the reviewerreviewer_id = tmp_info["reviewerID"]time_info = re.split(', | ', tmp_info["reviewTime"])review_time = time_info[2] + '-' + time_info[0] + '-' + time_info[1]# asin - ID of the productproduct_id = tmp_info["asin"]# overall - rating of the productrating = tmp_info["overall"]# if cc > 1000:#     break# print(tmp_info)# print(user_reviews)if product_id in games.keys():product_title = games[product_id]if reviewer_id in user_reviews.keys():user_reviews[reviewer_id].append((product_title, rating, review_time))else:user_reviews[reviewer_id] = [(product_title, rating, review_time)]if len(user_reviews) % 10000 == 0:print(f'Length of user_reviews: {len(user_reviews)}')cc += 1user_reviews_sorted = {}
for k, v in user_reviews.items():# 首先去重v = list(set(v))# 然后根据评价时间从小到大排序，表示用户的评价历史v_sorted = sorted(v, key=lambda x: x[2])# 选择具有7个及以上的评论样本if len(v) >= 7:# print(f'v: {v}, v_sorted: {v_sorted}')user_reviews_sorted[k] = v_sorted
print(f'Length of user_reviews_sorted: {len(user_reviews_sorted)}')

第三步 训练数据生成

# 总样本
samples = []
# 指令
instruction = "You are an assistant working on Video Games recommendations. Given the user's history of Video Games they have shopped, which includes the \"Title\" of the Video Games and the \"Rating\" the user rate (the Rating value is like or dislike), please decide whether the user likes to shop the target Video Games by outputting the order of their titles."
samples = []
cc = 0
for k, v in user_reviews_sorted.items():# print('-'*10)# print(v)sample_input = "User shopped Video Games histories (Title and Rating): \n"# 前面的当作对话历史for vv in v[0: -2]:# 当 rating 大于 3.0 的时候设置为 likeif vv[1] > 3.0:rating = 'like'# 当 rating 小于等于 3.0 的时候设置为 dislikeelse:rating = 'dislike'sample_input += "<Title: {}, Rating: {}>\n".format(vv[0], rating)sample_input += "Based on the Video Games histories, please sort the following two Video Games titles. The one in the front is what the user like and should be recommended to user: \n"# 最后两个设置为需要预测的目标sample_input += "<Title: " + v[-2][0] + '>\n'sample_input += "<Title: " + v[-1][0] + '>\n'# print(f'v[-1][1]: {v[-1][1]}, v[-2][1]: {v[-2][1]}')# 保证有一个是 like，有一个是 dislikeif (v[-1][1] > 3.0 and v[-2][1] <= 3.0) or (v[-1][1] <= 3.0 and v[-2][1] > 3.0):# print(f'v[-1][1] != v[-2][1]: {v[-1][1]}, {v[-2][1]}')if v[-1][1] > v[-2][1]:# likeoption1 = v[-1][0]# dislikeoption2 = v[-2][0]else:# likeoption1 = v[-2][0]# dislikeoption2 = v[-1][0]# chosen 是 like 在前面chosen = "<Title: " + option1 + '>\n' + "<Title: " + option2 + '>'# rejected 是 dislike 在前面rejected = "<Title: " + option2 + '>\n' + "<Title: " + option1 + '>'sample = {"instruction": instruction,"input": sample_input,"chosen": chosen,"rejected": rejected}# print(f'--------')# print(v)# print(sample)samples.append(sample)if len(samples) % 10000 == 0:print(f'Length of samples: {len(samples)}')# cc += 1# if cc > 10:#     breakprint(f'Length of samples: {len(samples)}')

第四步 划分 train 和 test 保存样本

# 首先打乱
shuffle(samples)train = samples[:int(len(samples)*0.8)]
test = samples[int(len(samples)*0.8):]print(f'总样本数: {len(samples)}，训练集样本数: {len(train)}，测试集样本数: {len(test)}')with open("./data/processed/rlhf_train.json", "w", encoding='utf-8') as save_file:json.dump(train, save_file, indent=4)with open("./data/processed/rlhf_test.json", "w", encoding='utf-8') as save_file:json.dump(test, save_file, indent=4) # , sort_keys=True

二、配置训练文档

rlhf_train.yaml

### model
model_name_or_path: /ZhipuAI/glm-4-9b-chat### method
stage: dpo
do_train: true
finetuning_type: lora
lora_target: all
lora_rank: 16
lora_alpha: 32
pref_beta: 0.1
pref_loss: orpo### dataset
dataset: amazon_video_games
template: glm4
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16### output
output_dir: ./saves/amazon_video_games_orpo
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 8
learning_rate: 5.0e-6
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000### eval
val_size: 0.1
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500

rlhf_inference.yaml

model_name_or_path: /ZhipuAI/glm-4-9b-chat
adapter_name_or_path: ./saves/amazon_video_games_orpo
template: glm4
finetuning_type: lora

三、模型预测

import json
from openai import OpenAI
from tqdm import tqdm# 加载模型
client = OpenAI(api_key="EMPTY",# 需要修改为大模型地址base_url="http://10.114.16.65:8000/v1/"
)

# 加载测试数据
test_file_path = "./data/processed/rlhf_test.json"
with open(test_file_path, "r", encoding='utf-8') as test_file:test_data = json.load(test_file)
print(len(test_data))

# 开始预测
labels = []
predictions = []
cc = 0
for each_test in tqdm(test_data):chat_completion = client.chat.completions.create(messages=[{"role": "system","content": each_test["instruction"]},{"role": "user","content": each_test["input"],}],model="glm4",)predictions.append(chat_completion.choices[0].message.content)labels.append(each_test["chosen"])if len(labels) % 100 == 0:correct = 0wrong = 0for l, p in zip(labels, predictions):l = l.strip()p = p.strip()# print(f'l: {l}, p: {p}')if l == p:correct += 1else:wrong += 1# print(f'\nl: {l}, \np: {p}')print(f'总样本数：{len(labels)}，准确数：{correct}, 错误数：{wrong}, 准确率：{correct / len(labels)}')cc += 1# if cc > 100:#     breakassert len(predictions) == len(labels)correct = 0
wrong = 0for l, p in zip(labels, predictions):l = l.strip()p = p.strip()if l == p:correct += 1else:wrong += 1
print(f'总样本数：{len(labels)}，准确数：{correct}, 错误数：{wrong}, 准确率：{correct/len(labels)}')