What is Vicuna?
Vicuna is an open-source research chatbot built by fine-tuning LLaMA on user-shared conversations collected from ShareGPT. The initial release introduced Vicuna-13B as a starting point for chatbot research and included a demo so people could examine its responses directly. The project was a collaboration among researchers from UC Berkeley, CMU, Stanford, UC San Diego, and MBZUAI.
What Vicuna can do
Vicuna handles multi-turn dialogue and was trained with a longer context window than Stanford Alpaca. In the release examples, Vicuna produced more detailed and better-structured answers than Alpaca. The project also published code for researchers who want to study how the chatbot was trained, served, and evaluated.
- Vicuna responds across multiple turns in a conversation.
- Vicuna supports a maximum context length of 2048 tokens.
- The Vicuna demo lets users test chatbot responses online.
- The Vicuna serving system can run multiple models across distributed GPU workers.
How Vicuna works
Vicuna was fine-tuned on approximately 70K conversations that users shared through ShareGPT. The team converted the collected HTML to Markdown, filtered out some low-quality or inappropriate examples, and split long conversations into segments that fit the model's context window. During fine-tuning, the training code computed loss only on the chatbot's replies rather than on the user messages.
Vicuna also has a distributed serving system that supports GPU workers from on-premises clusters and cloud environments. The team used this system to serve the online demo. The implementation combines a fault-tolerant controller with SkyPilot managed spot instances, which the team used to lower serving costs.
Who Vicuna is for
Vicuna is intended as a starting point for researchers and developers examining how language models behave in conversation. The online demo provides a way to inspect responses without setting up the serving system. The published weights and FastChat code are useful for those who want to look beyond the demo and study the training, evaluation, or serving implementation.
- Test multi-turn dialogue in the online demo.
- Study the fine-tuning method used for LLaMA.
- Compare chatbot replies using the published evaluation approach.
- Explore distributed GPU serving through FastChat.
Vicuna pricing and plans
Vicuna was released with public model weights, FastChat code, and an online demo intended for non-commercial research use. The release does not list paid plans or usage fees. The FastChat code is released under Apache License 2.0, while the demo is subject to the LLaMA model license, OpenAI's terms for generated data, and ShareGPT's privacy practices.
Strengths and limitations of Vicuna
Vicuna scored higher than LLaMA-13B and Alpaca-13B in the team's preliminary evaluation of 80 questions. The result reflects GPT-4's judgments of the answers, not an independently validated quality measure. The authors noted that GPT-4 was not a reliable judge of coding and mathematics and that a more rigorous evaluation was needed.
- Vicuna can struggle with reasoning and mathematics.
- Vicuna may give factually inaccurate answers.
- Vicuna was not sufficiently optimized to mitigate toxicity or bias.
- The training dataset was not released.
Getting started with Vicuna
Vicuna's release page links to the online demo, the Vicuna-13B model weights, and the FastChat repository. The demo was released for non-commercial experimentation with chatbot responses. FastChat contains the training, serving, and evaluation code for readers who want to inspect the implementation. The original ShareGPT training dataset is not part of the release.
Vicuna FAQ
What is Vicuna?
Vicuna is an open-source research chatbot built by fine-tuning LLaMA on approximately 70K user-shared conversations from ShareGPT. The Vicuna-13B release includes public model weights and code for training, serving, and evaluation.
Is Vicuna free?
Vicuna was released with public model weights, Apache License 2.0 code, and an online demo for non-commercial research use. The release page does not list paid plans or usage fees.
Can Vicuna run on on-premises GPUs?
Vicuna can be served with the project's distributed serving system, which supports GPU workers from both on-premises clusters and the cloud. The serving code is available in the FastChat repository.
How was Vicuna trained?
Vicuna was trained by fine-tuning LLaMA on approximately 70K conversations collected from ShareGPT. The team filtered some examples, split long conversations into shorter segments, and computed the fine-tuning loss only on the chatbot's replies.
How was Vicuna compared with ChatGPT?
Vicuna received 92% of ChatGPT's total score in a preliminary study in which GPT-4 judged answers to 80 questions. The authors described the method as non-scientific and said the results needed more rigorous evaluation.
From the community
No reviews yet
Be the first to review this tool. What you write helps the next person choose.




