One comments

Results 109 comments of

One

Installation issues

Could you please provide the full error log? It'd be more helpful.

function_call

Thank you for your suggestion! Currently in progress. Additionally, as a temporary solution, the current model can make function calls with appropriate prompts.

Any plan to train OpenChat on Mamba from Gu and Dao?

Thanks for your suggestion! We're actively working on new models and may consider Mamba, too.

Is OpenChat trained by supervised learning or reinforcement learning?

It's trained by C-RLFT, a fine-tuning method inspired by offline reinforcement learning. You can refer to our paper here for details. https://arxiv.org/abs/2309.11235

[Question] Can OpenChat be instruct-tuned for further downstream tasks?

OpenChat takes ShareGPT format data. You can convert the dataset for training using https://github.com/imoneoi/openchat/blob/master/ochat/data/generate_dataset.py BTW we will update instructions on custom data soon.

Getting "GPT4 Correct User" mid-response

Hi, have you changed the EOS token id to 32000? It seems to be a stop token issue.

is openchat 3.5 a multilingual model?

Yes. Are you using a quantized version?

is openchat 3.5 a multilingual model?

Can you paste your prompt here?

is openchat 3.5 a multilingual model?

Is the eos_token `` ? Btw the temperature should be lower (like 0.5) for Mistral models.

Why does the 8-GPUs experience a longer concurrent delay compared to the 2-GPUs?

Do you use tensor parallel for 8 GPUs? Tensor parallel will lead to a significant communication overhead. It's recommended to spin up separate server instances on each GPU and load...