Open Source ChatBots
- ChatGPT 2022/11/30. This was not open source but a challenge to the open source community.
- LLaMA 2023/02/23 Meta/Facebook. You need to fill out some paperwork and get approved to download this code. If you use a torrent and only want 13B be sure not to download the others as it gets huge.
- Alpaca 2023/03/13 Stanford. Created from LLaMA using training against ChatGPT. The training cost was about $600.
- GPT4ALL 2023/03 This has all the files needed to run and is open source. A bit over 4GB for the download.
- OpenChatKit 2023/03
- ColossalAI 2023/03 is another open source chatbot recently released.
- Dolly - 2023/03/24 trained on one machine in 30 minutes. Dolly was the name of a cloned sheep.
- Guanaco 2023/03
- Raven-RWKV-7B This uses RNN, or Recurrent Neural Network and not transformers like most LLMs.
- OpenAsisstant To release Apr 15, 2023. Demo
- Vicuna 2023/03 You can try Vicuna here. The paper on Vicuna explains that they trained on sample chat conversations found at ShareGPT. The training cost was about $300. This project is by students from UC Berkeley, CMU, Stanford, and UC San Diego. They used ChatGPT to rate answers from Vicuna against LLaMA and Alpaca and it rated Vicuna higher. Also using the name FastChat. Released the trained weight values as deltas from LLaMA weights on 2023/04/03. It can run in 60GB of CPU RAM or in 24 GB of GPU RAM.
- ChatGLM6b Using int4 can run on 6GB GPU video demo Chinese/English
- Cerebras-GPT-13B Download Announcement
- MiniGPT4 Enhancing Vision-language Understanding with Advanced Large Language Models demo
- StabilityLM 2023/04/20 Announcement demo
- BabyAGI Readme Twitter
- AutoGPT Readme Discord
- Jarvis Readme HuggingGPT: Solving AI Tasks with ChatGPT and
its Friends in Hugging Face - Red Pajama
- Hyena Hyena, a subquadratic drop-in replacement for attention
- HugginChat 2023/04/26 demo OpenAssistant/oasst-sft-6-llama-30b
- FastChat-V5 2023/04/28 fastchat-t5-3b-v1.0
- Stable-Vicuna-13b - 2023/04/28
- StarCoder-15B 2023/05/04 Good at writing programs
- MPT-7B 2023/05/05 Trained on MosaicML platform in 9.5 days ~$200k
- CodeT5+ 2023/05/16 best open source LLM at coding
This is very rapid activity in open source chatbots. Given the low training costs we expect many more chatbots to be released in the coming years with impressive performance progress. Open Source makes public each new idea and invention. It could evolve and progress very quickly.
In some sense AI is already being used to train AI, so The Singularity may be starting.
We will be trying out some of these chatbots. Our current test machine is a HP Envy with 16 core i9-12900K CPU, 128GB Ram and Nvidia GeForce RTX 3060. This was about $3,000 on Amazon when we bought it recently.
It seems running on your own computer is now "running locally" or "self hosted", as opposed to running in the cloud. We are interested in chatbots that can be run locally.
We downloaded and installed GPT4ALL. When running it only used about 4GB of RAM. It answered questions but was clearly not as good as GPT4 or Vicuna.
We downloaded Dolly and it was around 20GB so it could have all that is needed. Did not get it running yet.
Vicuna install is two posts after this.
Comments
Post a Comment