Vicuna

Introduced March 30, 2023 by LMSYS (the Berkeley/SkyLab team behind Chatbot Arena), Vicuna-13B was created by fine-tuning LLaMA on ~70K user-shared conversations from ShareGPT — at a training cost of roughly $300. Using GPT-4 as a judge, the team reported it reached over 90% of ChatGPT and Bard quality while beating LLaMA and Stanford Alpaca in 90%+ of comparisons.

Vicuna proved that a small team could produce a near-frontier chatbot on a shoestring budget, igniting the 2023 open-model fine-tuning boom and the local-LLM movement (it became the default model in oobabooga/text-generation-webui). The weights are available for non-commercial use only — the LLaMA base license restriction applies. A v1.5 refresh on Llama 2 followed in July 2023.