Image generation AI 'Stable Diffusion' developer releases chat AI 'StableVicuna'

Stability AI, the developer of Stable Diffusion, has announced the release of an open source chatbot AI `` StableVicuna ''. StableVicuna is a chatbot AI trained by further adjusting the chatbot AI ' Vicuna-13B ' based onLLaMA 13B , a large-scale language model developed by Meta.

Stability AI releases StableVicuna, the AI World's First Open Source RLHF LLM Chatbot — Stability AI
https://stability.ai/blog/stablevicuna-open-source-rlhf-chatbot

StableVicuna takes two approaches: 'prompt fine-tuning ' and 'reinforcement learning with human feedback (RLHF)'. In the past, fine-tuning instructions was a complicated task, so RLHF was almost never done. However, in recent years, Stablity AI says that StableVicuna has been realized because the RLHF dataset for chatbots has been provided as open source.

StableVicuna includes the OpenAssistant Conversations Dataset (OASST1) , a human-generated and human-annotated conversations dataset, GPT4All Prompt Generations, a dataset of over 430,000 prompts and responses generated with GPT-3.5 Turbo, Fine-tuning by Alpaca generated by OpenAI's text-davinci-003 engine. Furthermore, trlx is used for OASST1, Anthropic HH-RLHF , and Stanford Human Preferences for reinforcement learning and RLHF training.

As for what StableVicuna can do, Stability AI lists three things: 'can handle basic mathematics', 'can write code', and 'have grammar corrected'. At the time of writing the article, Stability AI says that the chat interface with StableVicuna has not been released and will be released soon.

StableVicuna is hosted on HuggingFace, a repository for AI, but only the weight difference is published, and in order to actually experience StableVicuna in a local environment, you need to be able to access LLaMA's original model. It seems that there is

CarperAI/stable-vicuna-13b-delta Hugging Face
https://huggingface.co/CarperAI/stable-vicuna-13b-delta