> For the complete documentation index, see [llms.txt](https://brindha.gitbook.io/mylearning/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://brindha.gitbook.io/mylearning/tools/vicuna.md).

# Vicuna

#### Vicuna: Open-Source Chatbot from LMSYS

**What is Vicuna?**

Vicuna is an open-source chatbot developed by a collaborative team from UC Berkeley, Carnegie Mellon University, Stanford University, and UC San Diego, operating under the Large Model Systems Organization (LMSYS). Released in March 2023, Vicuna was created by fine-tuning Meta's LLaMA base model on real user conversations — a key distinction from other instruction-tuned models of its era that relied on synthetically generated data.

***

**Origin and Motivation**

When Meta's LLaMA weights were leaked in early 2023, an ecosystem of fine-tuned derivatives rapidly emerged. Vicuna was one of the first and most influential of these, demonstrating that a fine-tuned model trained on authentic human-AI conversations could achieve remarkably high performance at very low cost. The entire training process for Vicuna-13B cost approximately $300 — a figure that became widely cited as evidence that competitive conversational AI no longer required massive corporate resources.

***

**Training Approach**

Vicuna was fine-tuned from LLaMA using supervised instruction fine-tuning on conversations collected from ShareGPT.com — a platform where users publicly shared their ChatGPT conversations. The initial version used approximately 70,000 conversations, and later versions such as v1.3 expanded this to around 125,000 conversations. The context length was extended from LLaMA's default of 512 tokens to 2,048 tokens to better handle multi-turn dialogue, and gradient checkpointing was used to manage the increased GPU memory requirements.

***

**Model Sizes and Versions**

Vicuna was released in 7B and 13B parameter sizes, both fine-tuned from their respective LLaMA counterparts. Multiple versions were released over time, with v1.1, v1.3, and v1.5 progressively improving performance and expanding context length. Later versions were fine-tuned from LLaMA 2 rather than the original LLaMA, benefiting from Meta's improvements to the base model. Vicuna v1.5 supported a 16K token context window using positional interpolation techniques.

***

**Performance**

At the time of release, preliminary evaluation using GPT-4 as a judge showed Vicuna-13B achieved over 90% of ChatGPT's quality and outperformed other open-source models including LLaMA and Stanford Alpaca in more than 90% of test cases. On GPT-4's scoring scale across 80 questions, Vicuna's total score reached 92% of ChatGPT's. While these evaluations were widely discussed, the team was transparent that this was a non-scientific evaluation methodology and that rigorous benchmarking remained an open challenge. This work contributed directly to the broader research field of LLM-as-a-judge evaluation methodology.

***

**FastChat Framework**

Alongside Vicuna, LMSYS released FastChat — an open-source framework for training, evaluating, and deploying instruction-following language models. FastChat provides command-line interfaces, OpenAI-compatible APIs, and web UI tools, making it possible for organizations to deploy Vicuna or train their own Vicuna-style models. FastChat also powers Chatbot Arena, LMSYS's crowdsourced evaluation platform where users blind-test different models and vote on responses.

***

**License**

Vicuna is available for non-commercial research use only. The non-commercial restriction inherited from LLaMA's original license limits its use in production commercial applications, which is a notable constraint compared to models released under Apache 2.0.

***

**Strengths**

* Trained on real human conversations rather than purely synthetic data, producing natural and well-structured responses
* Remarkably cost-efficient training demonstrating the power of fine-tuning at minimal expense
* Spawned the LLM-as-a-judge evaluation methodology, a lasting contribution to the field
* FastChat framework provides complete deployment tooling
* Strong multi-turn conversational capability at 7B and 13B scales
* Historically significant as one of the first high-quality open chatbot models

**Limitations**

* Non-commercial license restricts business use
* Based on older LLaMA generations — surpassed by newer base models and fine-tunes
* No multimodal capabilities
* Limited to English primarily
* No longer actively developed as a frontier model

***