Vicuna-13B

Vicuna-13B

by LMSYS

Open-source chatbot fine-tuned on LLaMA with 70K ShareGPT conversations

Open Source Neural Network API Python API Hugging Face
Visit Product
396 upvotes 1,474 views

About

Vicuna-13B is an open-source chatbot developed by researchers at UC Berkeley, CMU, Stanford, and UCSD. It was created by fine-tuning Meta's LLaMA model on approximately 70,000 user-shared ChatGPT conversations from ShareGPT, resulting in a model that exhibited surprisingly strong instruction-following capabilities at a fraction of the cost and compute of training from scratch.

When evaluated by GPT-4 in blind comparisons, Vicuna-13B achieved over 90% quality relative to ChatGPT and outperformed other open-source models like LLaMA and Alpaca on complex, multi-turn conversations. This groundbreaking result demonstrated that fine-tuning on high-quality conversation data could dramatically close the gap between open-source and closed-source models.

Vicuna played a crucial role in the open-source AI community by proving that smaller academic groups could create competitive conversational AI through clever data curation rather than massive compute budgets. It spawned an entire generation of fine-tuned LLaMA variants and established conversation data quality as a key variable in model capability.

Product Features

- Fine-tuned on 70K high-quality ShareGPT conversations
- 13B and 7B model sizes
- Strong multi-turn conversation ability
- Open weights for research and commercial use
- Compatible with all LLaMA inference frameworks
- Runs on consumer GPUs (13B requires 28GB VRAM)
- Quantized versions for reduced memory footprint
- Available on Hugging Face model hub
- FastChat integration for easy deployment
- Basis for hundreds of derivative fine-tuned models

About the Publisher

Vicuna was created by the LMSYS (Large Model Systems Organization) group, a collaboration between researchers at UC Berkeley, CMU, Stanford, and UC San Diego. LMSYS also developed Chatbot Arena — a crowdsourced evaluation platform for comparing LLMs through human preference judgments — and the FastChat framework for serving language models. The group has made significant contributions to open-source AI through model releases, evaluation methodology, and inference optimization research.