Vicuna: The Model That Proved
Fine-Tuning Works

Why Vicuna Mattered: A Brief History

To understand Vicuna, you need to understand what people believed in early 2023. The prevailing assumption was that truly capable chat models required vast proprietary datasets, enormous compute budgets, and specialised RLHF pipelines that only OpenAI, Anthropic, and Google could afford. Then, in a single month, three teams shattered that assumption in sequence.

How Vicuna Was Built

Vicuna's training recipe was simple enough to fit in a blog post and reproduce on a weekend — which is exactly why it was so influential. The key innovations over Alpaca were in data quality, not training technique.

Training data: ShareGPT conversations

ShareGPT was a website where ChatGPT users publicly shared their conversations. The LMSYS team scraped approximately 70,000 of these conversations. Unlike Alpaca's single-turn synthetic instructions, ShareGPT conversations are:

• Multi-turn — real back-and-forth dialogue spanning multiple exchanges
• User-initiated — prompts reflect what real users actually want, not what researchers think they want
• Diverse — covering coding, writing, Q&A, roleplay, summarisation, and everything in between
• Long-form — many conversations include detailed, structured responses that teach format as well as content

Training modifications

LMSYS made several technical improvements over the basic instruction-tuning approach used in Alpaca:

• Gradient checkpointing — reduces memory usage during training, allowing longer sequences to fit in GPU memory
• Loss computed only on model responses — the training loss is masked on user-turn tokens; the model only learns to predict the assistant's side of the conversation. This was a key fix over naive approaches that trained on all tokens equally
• Sequence length expansion — ShareGPT conversations are often long; Vicuna's training pipeline handles sequences up to 2,048 tokens (later extended in v1.5)
• SponsorBlock-style data cleaning — conversations with hallucinated context, heavily truncated content, or other quality issues were filtered

A chat between a curious user and an artificial intelligence assistant.
The assistant gives helpful, detailed, and polite answers to the user's questions.

USER: {user_message}
ASSISTANT: {assistant_response}</s>
USER: {next_user_message}
ASSISTANT:

Vicuna Versions: Which to Use

Where Vicuna Stands Today

Vicuna-1.5 was competitive at release in July 2023. By late 2023, the field had moved significantly. The table below contextualises where Vicuna stands relative to its contemporaries and successors.

The Licence Problem

Running Vicuna Locally

Ollama

# Vicuna 7B v1.5 (fastest, lowest memory)
ollama run vicuna

# Vicuna 13B v1.5 (higher quality, needs 10GB+ RAM)
ollama run vicuna:13b

# Vicuna 33B (research; needs 24GB+ RAM)
ollama run vicuna:33b

Hardware requirements (v1.5)

See deployment quickstart for platform-specific Ollama installation on macOS, Linux, and Windows.

Python example

import ollama

response = ollama.chat(
    model='vicuna:13b',
    messages=[
        {
            'role': 'user',
            'content': 'What are the main differences between REST and GraphQL APIs?'
        }
    ]
)

print(response['message']['content'])

# For research/personal use only — CC BY NC 4.0 licence
# For production: use openchat or starling-lm instead

Honest Assessment

When to Choose Vicuna