LLaMA's consequential projects.
- llama.cpp by Georgi Gerganov, a port of Facebook's LLaMA model in C/C++
- bloomz.cpp, a port of BLOOM built on top of llama.cpp
- gpt4all, an assistant-style large language model with ~800k GPT-3.5-Turbo Generations based on LLaMa
- Dalai by @cocktailpeanut, a single command installer to run LLaMa locally with a web interface and API
- later expanded to include alpaca.cpp
- Stanford Alpaca, a model fine-tuned from the LLaMA 7B model on 52K instruction-following demonstrations to function like ChatGPT
- alpaca-lora, a recreation of Stanford Alpaca, based on LLaMA 7B and the Stanford instruction training set using Low-Rank Adaptation instruct tuning
- alpaca.cpp, a fork of llama.cpp that uses alpaca-lora
- many children of this who have followed the method for the larger LLaMA models, such as chansung/alpaca-lora-13B and chansung/alpaca-lora-30B
- alpaca-native, a recreation of Standford Alpaca, based on LLaMA 7B and the Stanford instruction training set using native fine-tuning (not LoRA)
- alpaca-native-4bit, a 4-bit quantisation of alpaca-native made with GPTQ-for-LLaMA (which uses GPTQ)
- Databricks Dolly, a model fine-tuned from GPT-J using the Alpaca training set, demonstrating surprisingly high quality instruction following behavior not characteristic of GPT-J
- alpaca-lora, a recreation of Stanford Alpaca, based on LLaMA 7B and the Stanford instruction training set using Low-Rank Adaptation instruct tuning
- Vicuna-13B, a model fine-tuned from LLaMA-13B based on 70k conversations from ShareGPT, with a training method improving upon Alpaca's
Links
- AGI Sphere's writeup on the LLaMa family, as of April 2023