search

kaistAI / SelFee

Official codebase for "SelFee: Iterative Self-Revising LLM Empowered by Self-Feedback Generation"
Apache License 2.0
222 stars 20 forks source link

uncensored version? #5

Closed EdwardDali closed 1 year ago

EdwardDali commented 1 year ago

Hi, Realy nice project. Any plan on releasing an uncensored version? Usually they have better performance. Also a wiki with use-cases and how-to would be nice. Also, an increase of the context window seems possible, see the latest projects released by theBloke on huggingface/ kaiokendev work.

Kudos,

seonghyeonye commented 1 year ago

Hi, sorry for the late response. Could you specify what you mean by uncensored version? We also agree that the increase of the context window using other techniques would lead to better performance since our work leads to increased sequence length. Thank you.

EdwardDali commented 1 year ago

Hi, thanks for the reply.

Looking forward for your next model :) An uncensored model is practicaly a model for which responses such as "as an AI language model..." are removed. So for instance responses that contained alignment / moralizing are removed and the model is actually responding and not censoring some responses. See for instance uncensored models on hugginface such as: https://huggingface.co/ehartford/Wizard-Vicuna-13B-Uncensored or datasets such as https://huggingface.co/datasets?sort=trending&search=uncensored

Related to the context size a recent model seem to have a 32K context window: https://huggingface.co/togethercomputer/Llama-2-7B-32K-Instruct

seonghyeonye commented 1 year ago

Thank you for your suggestion! We will definitely try to explore those directions :)