Jiaxin Shan

Is there a way to evaluate performance between alpaca and alpaca-lora?

7

I am just curious whether there's a scientific way to compare the performance between alpaca and alpaca -lora? Does the community have some evaluation scripts to run?

Using lora to finetune domain specific data?

3

Hi, I know this question may not be directly related to the project. Just try to get some clues. I do have some wikis, dialogs and I'd like to change...

How to convert dialog data to instructions for finetuning?

3

Hi community, I am trying to fintune a chatbot with domain specific knowledges using alpaca way for research purpose. I do have some wikis, dialogs. does anyone know how to...

Unable to save the mode weights - GPU OOM

9

I am finetuning vicuna using 4 * A100-80G GPUs. I meet some problem after finish training, ``` {'loss': 1.3641, 'learning_rate': 4.815273327803183e-08, 'epoch': 0.97} {'loss': 1.35, 'learning_rate': 2.7095433213097933e-08, 'epoch': 0.97} {'loss':...

The hugging face docker image is invalid and failed to start

3

``` docker run -it -p 7860:7860 --platform=linux/amd64 registry.hf.space/microsoft-hugginggpt:latest python app.py Unable to find image 'registry.hf.space/microsoft-hugginggpt:latest' locally latest: Pulling from microsoft-hugginggpt fb668870d8a7: Pull complete 8a612414e2bc: Pull complete 2c12f5dee74d: Pull complete e8b64516db7f:...

Could I know how do you evaluate your model performance?

3

See title. What's the dataset? Did you run any evaluation steps?

Encounter the runtime error training with lora and flash_attention together

2

I am using fschat to finetune a vicuna model with lora and get following error. I tried normal trining with flash attn as well and it works fine. It's probably...

Address https://github.com/lm-sys/FastChat/issues/350 Verification: controller ``` 2023-05-05 00:51:45 | INFO | gradio_web_server | ==== request ==== {'model': 'vicuna-13b', 'prompt': "A chat between a curious user and an artificial intelligence assistant. The...

Address https://github.com/lm-sys/FastChat/issues/847 Test procedure: 1. Start controller, model worker1, model worker2 and web, model list does show 2 models 2. Kill model worker2 and check dropdown list again and we...

I notice we do plan to support reload mode, https://github.com/lm-sys/FastChat/blob/a94fd259a97128f7f4483ddb760690f467888d84/fastchat/serve/gradio_web_server.py#L533-L535 Seems this is not implemented yet https://github.com/lm-sys/FastChat/blob/a94fd259a97128f7f4483ddb760690f467888d84/fastchat/serve/gradio_web_server.py#L506-L522 I will add a PR to support this feature