What it does
Vicuna is a chat model from the LMSYS team, a joint effort of researchers at several universities. The 13B version was made by fine-tuning Meta's LLaMA on about 70,000 conversations that people shared on ShareGPT, using training code adapted from Stanford Alpaca to handle multi-turn chats and longer context. The team released the Vicuna-13B weights along with training, serving and evaluation code in the FastChat repository, and ran an online demo. Their informal evaluation used GPT-4 as a judge and put Vicuna at roughly 90 percent of ChatGPT quality, but they stress that this was not a rigorous test. The announcement dates from March 2023. The weights and demo are for non-commercial use under the LLaMA license and the terms of the training data, while the code is Apache 2.0. The team lists weak spots in reasoning, math, factual accuracy and safety.
Capabilities
- Chat model fine-tuned from LLaMA on about 70,000 shared conversations
- Model weights plus training, serving and evaluation code released
- Context length extended to 2,048 tokens for long conversations
- Serving system that can run multiple models across distributed GPU workers
When to use it
- Studying how an early open chat model was trained, served and evaluated
- Experimenting with an open LLaMA-based chat model on your own hardware for non-commercial work
- Using it as a baseline in open-model research
Alternatives to explore
- Stanford Alpaca
- Koala
