Text Generation
Transformers
English
llama
Inference Endpoints

USE 1.1 vicuna since it no longer has the issue of feedback / talkijng to its self?

#10
by nucleardiffusion - opened

vicuna-13b-free-q4_0.bin sufferes from these issues but the newer vicuna 1.1 does not

https://huggingface.co./eachadea/ggml-vicuna-13b-1.1

Then plan is to move on to 1.1 once the V4 of the unfiltered dataset is ready. Hopefully the improved separator in 1.1 will resolve the whatever issues the 1.0 model has regarding this... Unfortunately the training on 1.0 had already started when the 1.1 was released soon after. But not a huge deal, it was expected that we have to iterate on the dataset in any case.

I'm using "### Human", "### Assistant" in Custom stopping strings in oobabooga/text-generation-webui to remove such issue like talking to itself

Sign up or log in to comment