[中文|English]
使用朴素贝叶斯思想来扩展LLM的Context处理长度。
现在,任何LLM都可以利用NBCE成为可以处理任意长Context的模型了(只要算力足够)!
基于朴素贝叶斯所启发的公式:
细节请看博客:https://kexue.fm/archives/9617
所给的Demo包含12段不同的Context,总长度为9000多字,连同8个问题一次性输入到模型中(测试模型训练长度为2048,参数量为7B,可以在OpenBuddy下载),模型能够逐一根据所给Context正确回答这8个问题。值得指出的是,所有的Context、问题和答案加起来,超过了1万字!另外,有朋友简单尝试了简历匹配和作文打分应用,效果也尚可,非常建议大家亲自调试一下。
最新测试结果:在8*A800下,7B模型可以处理50k的context,并能正确地做阅读理解。(没有用完所有GPU,大概消耗160G显存)
- 即插即用
- 模型无关
- 不用微调
- 线性效率
- 实现简单
- 效果尚可
- 可解释性
@inproceedings{nbce_naacl,
author = {Jianlin Su and
Murtadha Ahmed and
Bo Wen and
Luo Ao and
Mingren Zhu and
Yunfeng Liu},
editor = {Kevin Duh and
Helena G{\'{o}}mez{-}Adorno and
Steven Bethard},
title = {Naive Bayes-based Context Extension for Large Language Models},
booktitle = {Proceedings of the 2024 Conference of the North American Chapter of
the Association for Computational Linguistics: Human Language Technologies
(Volume 1: Long Papers), {NAACL} 2024, Mexico City, Mexico, June 16-21,
2024},
pages = {7791--7807},
publisher = {Association for Computational Linguistics},
year = {2024},
url = {https://doi.org/10.18653/v1/2024.naacl-long.431},
doi = {10.18653/V1/2024.NAACL-LONG.431},
timestamp = {Thu, 29 Aug 2024 17:13:57 +0200},
}
QQ交流群:808623966,微信群请加机器人微信号spaces_ac_cn