ColossalAI
ColossalAI copied to clipboard
hpcaitech
[DOC]: optimize readme of ChatGPT
📚 The doc issue
A new user is hard to start training after reading the readme.
I list some FAQs which users may concern:
- How to prepare training dataset? What does the dataset look like?
- How to save models?
- How to use the fine-tuned models to generate?
- How to use LoRA? How does it affect training?
The biggest question in my mind is: How to initialize optimizer states when finetuning a pre-trained model?
Hi, I found this picture a little confusing. I think if there is some description, it would be much better.
Hi, I found this picture a little confusing. I think if there is some description, it would be much better.
Yes, at least tell the readers: the green part is the trainable parameters and the rest parts are frozen parameters.
I think we need a detail blog to tell user what is reward model, critic model, sft-model
I think instructs LM is better than chatgpt, chatgpt is a serving, instruct LM is an algorithm
This picture is from InstructGPT, maybe there should be copyright information?
Thanks for your remind, and we mention the reference here.
##Todo
- implement PPO training and fine-tuning
We should also mentioned PPO fine-tuning. This looks a lot like we haven't implemented finetune function yet, because we plan to implement PPO-ptx fine-tuning.
Clear error message or doc to remind user 'pip install .'
TODO ckpt Seriously affect the actual use of the user, need to be fixed as a priority
Hi, I found this picture a little confusing. I think if there is some description, it would be much better.
Yes, at least tell the readers: the green part is the trainable parameters and the rest parts are frozen parameters.
I've already completed a detailed doc to explain about the training process, it will be released soon.
We have updated a lot. Please check https://github.com/hpcaitech/ColossalAI/tree/main/applications/Chat This issue was closed due to inactivity. Thanks.