Facico / Chinese-Vicuna

Chinese-Vicuna: A Chinese Instruction-following LLaMA-based Model —— 一个中文低资源的llama+lora方案,结构参考alpaca
https://github.com/Facico/Chinese-Vicuna
Apache License 2.0
4.14k stars 422 forks source link

至死都无法学到三国相关的内容 #201

Open lucasjinreal opened 1 year ago

lucasjinreal commented 1 year ago

我看数据集里面有许多关于三国的内容,例如诸葛亮怎么死的,赵云怎么死的。

我测试下来,至死都没有学到任何三国内容。(epoch 3)

请问何故?

>诸葛亮是谁
诸葛亮(Sho Kakuryu)是一名欧州棋士,目前在英国住戶。他在2012年代排名前十名,在2014年代排名前十名,2016年代排名前
>what is the capital of America
The capital of the United States of America is Washington, DC.
>When is the Thanks Giving Day 
Thanksgiving Day is celebrated in the United States on the fourth Thursday of November.
>Who is the 45th president of America
Donald Trump
>who is the 44th president of America
Barack Obama
>How do u think of him
I think of him as a kind and loving person who has a giving spirit. He is always willing to help out and is a great person to talk to.
>解释给定成语所隐含的概念。
“寻找”的概念涉及寻找或找到一个特定的东西。
>诸葛亮是如何死的?
猫是由诸葛亮战胜于毒毒蛇的作者,他是毒毒蛇攻击的一名批员。诸葛亮是为了找到毒毒蛇的追随而逃跑的,但他努力的追随
>赵云是如何死的?
赵云是终身的一位革命性思想家和宪法学家,他在他的独特的思想方面遍及了各种领域,包括政治、经济、社
>

感觉英文能力得到了保留,但是中文能力的确是比较差。

想请教:

  1. 正常与否
  2. 如何提升
Facico commented 1 year ago

可能是正常的,因为在70w的数据量上3个epoch确实不可能把数据集的全部知识学会(而且用的是lora,吃不了多少新的知识)。这些知识最好是在预训练的时候就塞进去,否则只能专门 微调(比如用continue finetune的代码)

lucasjinreal commented 1 year ago

image

现在tiao调的baichuan还可以

Facico commented 1 year ago

嗯,这个中文基底可以的,相当于中文llama了