search

project-baize / baize-chatbot

Let ChatGPT teach your own chatbot in hours with a single GPU!
https://arxiv.org/abs/2304.01196
GNU General Public License v3.0
3.16k stars 282 forks source link

Performance cannot align with demo #28

Closed LIO-H-ZEN closed 1 year ago

LIO-H-ZEN commented 1 year ago

base_model=decapoda-research/llama-7b-hf lora_model=project-baize/baize-lora-7B python app.py $base_model $lora_model

same Top_p, Temperature

even 13b performs poor

guoday commented 1 year ago

Which two models did you compare?

JetRunner commented 1 year ago

The code used by the online demo is here: https://huggingface.co/spaces/project-baize/Baize-7B/blob/main/app.py Which is exactly the same as local demo. Can you provide more details?

LIO-H-ZEN commented 1 year ago

I overlooked the randomness, my fault... I turn the Temperature to 0.1 and the Top_p to 1.0. This time the local result is exactly the same as the online result. Besides, as 13B seems similar to 7B, i try to test 30B, but encounter this problem: https://github.com/project-baize/baize-chatbot/issues/31