模型介绍
- 名称：用户情感识别器
- 简介：随着互联网的不断发展，公司急需获取用户对其产品的反馈。然而，人类语言常常涉及讽刺、反语等复杂现象，这使得准确把握用户情感变得困难。该模型的设计目标是在尽可能有效地解读客户反馈的基础上，准确识别客户的情感。随着对用户见解的需求不断增加，该模型旨在通过处理语言的复杂性，提供对客户情感的细致理解，以弥补这一差距。
实验环境：
- 8核 32GB 显存24G
- ubuntu22.04-cuda12.1.0-py310-torch2.1.2-tf2.14.0-1.11.0
训练方法：
- 数据集：jd-sentiment-zh
- 微调的模型：qwen-1_8b
- 超参数：sfttype='lora', traindataset_sample=2000
示例代码
import os
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
import torch
//train
from swift.llm import (
DatasetName, InferArguments, ModelType, SftArguments,
infer_main, sft_main, app_ui_main, merge_lora_main
)
model_type = ModelType.qwen_1_8b
sft_args = SftArguments(
model_type=model_type,
sft_type='lora',
train_dataset_sample=2000,
dataset=[DatasetName.jd_sentiment_zh],output_dir='output')
result = sft_main(sft_args)
best_model_checkpoint = result['best_model_checkpoint']
print(f'best_model_checkpoint: {best_model_checkpoint}')
//infer
torch.cuda.empty_cache()
infer_args = InferArguments(
ckpt_dir=best_model_checkpoint,
load_dataset_config=True,
do_sample=False)
result = infer_main(infer_args)
推理效果
{"response": "positive", "query": "Task: Sentiment Classification\nSentence: 擦玻璃很好、就是太小了\nCategory: negative, positive\nOutput: ", "label": "positive"}
{"response": "negative", "query": "Task: Sentiment Classification\nSentence: 店家太不负责任了，衣服质量太差劲了，和图片上的不一样\nCategory: negative, positive\nOutput: ", "label": "negative"}
{"response": "positive", "query": "Task: Sentiment Classification\nSentence: 送国际友人挺好的，不错不错！\nCategory: negative, positive\nOutput: ", "label": "positive"}
{"response": "positive", "query": "Task: Sentiment Classification\nSentence: 很好,装好一定很漂亮\nCategory: negative, positive\nOutput: ", "label": "positive"}
{"response": "negative", "query": "Task: Sentiment Classification\nSentence: 东西给你退回去了，你要黑我钱！！！\nCategory: negative, positive\nOutput: ", "label": "negative"}
{"response": "positive", "query": "Task: Sentiment Classification\nSentence: 送货很快，书是正品，买书一直京东是首选！\nCategory: negative, positive\nOutput: ", "label": "positive"}
{"response": "positive", "query": "Task: Sentiment Classification\nSentence: 口感相当的好 都想买第二次了\nCategory: negative, positive\nOutput: ", "label": "positive"}
{"response": "negative", "query": "Task: Sentiment Classification\nSentence: 硅胶味道太重，样子与图片差距太大\nCategory: negative, positive\nOutput: ", "label": "negative"}
{"response": "negative", "query": "Task: Sentiment Classification\nSentence: 很伤心，买了放到三星n4尽然不能用，客服各种推\nCategory: negative, positive\nOutput: ", "label": "negative"}
{"response": "positive", "query": "Task: Sentiment Classification\nSentence: 质量不错，大小合适，应当是正品！但是我买的是黑灰，发来的却是纯黑，懒得换了，给个差评，希望以后改进！\nCategory: negative, positive\nOutput: ", "label": "negative"}
评论