Stanford Alpaca
Stanford's 2023 LLaMA fine-tune that made cheap instruction tuning standard
Overview
Stanford Alpaca is a March 2023 research model from Stanford CRFM: Meta's LLaMA 7B fine-tuned on 52,000 instructions generated with OpenAI's text-davinci-003, for under $600. Its demo came down within about a week, and what remains is the dataset, code and a weight diff, all restricted to non-commercial research use.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 14 Oct 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
Use cases
What you can produce with Stanford Alpaca
- The 52,000-example instruction dataset generated with text-davinci-003 (CC BY-NC 4.0)
- The data-generation code implementing Stanford's modified self-instruct method
- Fine-tuning scripts and hyperparameters for LLaMA 7B and 13B
- A weight diff on Hugging Face that rebuilds Alpaca 7B from Meta's original LLaMA weights
- The Alpaca prompt template used for instruction-and-input prompts
- A documented 2023 reference point for research on synthetic instruction data, for non-commercial use
Alpaca showed in March 2023 that LLaMA 7B, fine-tuned on 52,000 GPT-generated instructions for under $600, could follow instructions much like text-davinci-003 in its authors' blind test. The dataset, the training code and the self-instruct generation pipeline are still public and still a clear way to learn how instruction tuning works.
It is not deployable. The demo has been offline since March 2023, rebuilding the model needs Meta's original LLaMA weights plus a weight diff, and the data and diff are CC BY-NC with Stanford stating commercial use is prohibited. Study it; build on something newer.
What we found
Stanford Alpaca is a research model that Stanford's Center for Research on Foundation Models published on 13 March 2023. Eight authors, among them Percy Liang and Tatsunori Hashimoto, fine-tuned Meta's LLaMA 7B on 52,000 instruction-and-answer pairs generated by OpenAI's text-davinci-003 using a modified self-instruct method. Their post puts the data at under $500 of OpenAI API calls and the training at three hours on eight 80GB A100 GPUs, under $100 on most clouds. In a blind comparison rated by the five student authors, Alpaca won 90 contests to text-davinci-003's 89. That was the result that mattered: a small model, cheaply tuned on another model's outputs, could follow instructions much like a commercial one.
The live demo did not last. It went up with the announcement and was offline by 21 March 2023, when The Register reported the takedown; Stanford's post says the team brought it down because of hosting costs and the inadequacy of its content filters. The repository still says the demo is suspended until further notice.
What remains is the recipe. The GitHub repository holds the 52K dataset, the data-generation and fine-tuning code, and instructions for rebuilding Alpaca 7B from a weight diff on Hugging Face, which only works if you already hold Meta's original LLaMA weights. The repository is not archived, but its last commit to the main branch was in May 2023, and it names no successor project.
The licensing rules out products. The code is Apache 2.0, but the dataset and the weight diff are CC BY-NC 4.0, and Stanford's post says any commercial use is prohibited, for three reasons: LLaMA's non-commercial licence, OpenAI's terms against building competing models from its outputs, and the absence of adequate safety measures. Read it to understand how cheap instruction tuning began; for anything you plan to ship, use a current, properly licensed open model.
Compiled by RECATOOLS from public sources: vendor documentation and pricing, code repositories and published reporting.
Notable facts
- Stanford Alpaca was trained for $600 — approximately 150,000 times cheaper than the estimated cost of GPT-3's training.
- The research demonstrated that a 7B model could approach ChatGPT quality on many tasks, fundamentally changing beliefs about the minimum size required for useful instruction following.
- Alpaca's self-instruct methodology — using GPT-4 to generate training data — is now used to train virtually every open-source instruction-tuned model released since 2023.
Frequently asked questions
About this listing
This entry was compiled from publicly available data including Stanford Alpaca's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with Stanford Alpaca unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to Stanford Alpaca directly →
Spotted something out of date? Suggest an update →
Related reading
Health & Wellness
Fine-Particle Air Pollution Linked to Higher Hospital Admission Risk for Up to 21 Days, St...
Health & Wellness
A Guide for New Mothers Missed Its Main Goal but Nearly Doubled Primary Care Visits
AI & ML
Google Stopped Paying for One Kind of Open-Source Bug Report After a Flood of Invalid Ones
Alternatives to Stanford Alpaca
More in LLMs & Chat
