Vicuna is an open-source conversational model fine-tuned from Llama 2 for virtual-assistant tasks. Its practical appeal is efficiency: the 13-billion-parameter model is reported to approach ChatGPT’s accuracy while requiring fewer resources than larger alternatives. section 10.3.13
Its development and deployment emphasize:
- Training on roughly 90,000 ShareGPT conversations to adapt a foundation model to dialogue. section 10.3.13
- Comparing chatbots through Hugging Face and the LMSYS arena. section 10.3.13
- Improving domain accuracy by combining open models, fine-tuning, semantic search, and retrieval-augmented generation. section 10.3.13
A useful limitation to remember
When asked whether it can reason about physical things, a Vicuna response states that it generates text but cannot directly perceive, sense, or interact with the physical world.
