generate_question
1.0.0
This project attempts to generate related tasks based on language models. The first version uses million-level questions to fine-tune the T5 model, and the code is in the ./t5_tuning directory. The second version attempts to generate fine-tune based on gpt-2. This update is based on ChatGLM-6b+lora fine-tune the first version of the model, and adds a new type of fine-tune task.
(1) Updated the first version of the model based on ChatGLM-6b+lora (2) Newly generated propt data, new propt learning tasks (4) The query library has been sorted out to reach 40 million, and cleaned it
(1) Keyword to question generation (2) Keyword generation (3) The data set for fine-tuning similar question generation is in the ./chatglm_tuning/data directory, about 100,000+;
(1) On the propt_data dataset, 3 epochs were trained;
(2) Training loss, from the initial 3.0 to about 1.3; after about 2epoch, the loss drop has been particularly slow;
(3) For detailed training parameter settings, see the code;
#Keyword control generation
prompt:帮我生成10条包含[流量,套餐]关键词的问句
answer:
1.移动流量包怎么取消流量套餐
2.手机流量套餐有哪些?价格如何查询呢
3.这个有什么的流量套餐吗?可以换卡了吗
4.有免费的流量套餐吗
5.有没有免费流量套餐推荐一下啊!求大神们解答
6.联通流量超出后需要购买什么流量套餐
7.现在哪个网络套餐送更多流量
8.流量多少合适 选哪种套餐好点
9.电信流量套餐好吗
10.移动流量多吗?买套餐怎么样划算呀!谢谢各位大大回答!!
#Query keyword generation
prompt:[美国CPI为何暴涨]问句包含哪些关键词
answer:
包含的关键词有: 美国,CPI
#Similar Question Generation
prompt:帮我生成9条与[炒股软件哪个好用]相似的问句
answer:
1.请问有什么好的炒股工具可以推荐吗
2.想要在股市上获得成功,应该选择哪个炒股软件呢?请帮忙解答一下!
3.我想知道有哪些好用的炒股应用程序值得大家使用
4.股市新手应该如何选择适合自己的炒股软件呢
5.能否介绍一下哪些炒股软件适合股票投资者
6.我想学习更多有关股市的知识和技能,有没有好的炒股软件可以参考下的呢
7.现在有很多不同的炒股软件可供选择,不知道哪种最实用、最好用
8.对于刚刚入门的人来说,用什么炒股软件比较合适比较好
9.想请教各位高手们,哪些炒股软件是投资必备的好帮手啊!
(1) Judging from the training effect of this version, it feels that the training based on chatglm+lora is not sufficient, and the fitting degree of propt_data is relatively low. I am not sure whether to increase the round or adjust the lora parameters;
(2) When doing the task of generating keywords for question-based keywords, many important words are not recognized, which is related to the low quality of propt data;
(1) Continue to optimize the chatglm+lora model;
(2) Add or optimize the prompt task;
(3) Consider how to apply the query library, or build a knowledge base, and combine the model with the knowledge base to do query-related generation tasks.
