News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
Stanford Alpaca

Stanford Alpaca

A cheap instruction-tuned LLaMA 7B built at Stanford in 2023

Stanford Alpaca is an instruction-following large language model made by researchers at Stanford University in March 2023. It was built by fine-tuning LLaMA 7B on 52,000 demonstrations produced with OpenAI's text-davinci-003, and it drew notice for showing that a capable instruction-following model could be made cheaply from an existing foundation model.

Share on WhatsApp

  • Fine-tuned from LLaMA 7B in March 2023
  • Trained on 52,000 demonstrations grown from 175 human seeds
  • Data generation cost under US$500; fine-tuning under US$100
  • About three hours on eight 80 GB Nvidia A100 GPUs
  • Weights and dataset released for research only, not commercial use

What it is

Stanford Alpaca is a language model that follows written instructions, produced at Stanford University and introduced in March 2023. It starts from LLaMA 7B, a base model with seven billion parameters, and is adapted to respond to commands rather than simply continue text. The team built it as an inexpensive research tool for studying how instruction-following behaviour emerges. Its training set of 52,000 prompt-and-response pairs was generated with OpenAI's text-davinci-003, and the whole effort cost only a few hundred dollars.

How it came about

The work began as a low-budget research project on instruction-following systems. Starting from 175 examples written by people, the researchers used a modified form of the Self-Instruct approach to expand the set to 52,000 demonstrations, all produced through text-davinci-003. Creating that data cost under US$500. Stanford published the dataset and training code on GitHub for research use, and briefly ran a public demo before taking it down because of hosting expense and weak content filtering.

Inside the model

The model is LLaMA 7B after fine-tuning on 52,000 instruction-and-answer pairs. Those pairs were grown from 175 human-written seeds using a variant of Self-Instruct, with text-davinci-003 supplying the outputs. Fine-tuning ran for roughly three hours across eight Nvidia A100 cards with 80 GB each, an outlay estimated below US$100. Because the base weights come from LLaMA and the training text came through OpenAI's API, the result could not be released for commercial purposes.

What it can do

In a small preliminary comparison, Alpaca was preferred 90 times against 89 for text-davinci-003, so the two were judged about even. The researchers stressed that this test was limited and did not settle the question of quality. They also noted that the model carries over weaknesses from LLaMA and from the data used to train it, including hallucinations, false information, toxic wording and stereotyped answers.

How to use it

Stanford put the dataset and the training code on the project's GitHub repository for research use. The code is under Apache 2.0, while both the dataset and the weight diff carry CC BY-NC 4.0, which blocks commercial use. That restriction follows from the LLaMA base and from data made with OpenAI's API. A public demo existed briefly and was then withdrawn.

Safety and testing

Content filtering in the demo was judged inadequate, and hosting costs were high, so the public showcase was pulled shortly after launch. The team warned that the system can hallucinate, spread misinformation, emit toxic language and repeat stereotypes. It was never meant for general deployment.

Reception and debate

The project drew notice for demonstrating that an instruction-following model could be built from an existing foundation model at very low cost. Its own authors kept expectations modest, describing the evaluation as preliminary and pointing out inherited flaws. Commercial release was ruled out because of the underlying LLaMA licence and the OpenAI-generated training data.

Why it matters

Alpaca showed that instruction tuning of a mid-sized open base model could be done for a few hundred dollars, making such experiments accessible to small teams. It sits alongside other LLaMA-derived research systems, such as Vicuna, that followed the same pattern of adapting a released base model to conversational use.

Questions people ask about Stanford Alpaca

What is Stanford Alpaca?

It is an instruction-following large language model created at Stanford University in March 2023. Researchers took LLaMA 7B and fine-tuned it on 52,000 instruction-and-answer pairs so it would respond to commands rather than just continue text.

How was it trained?

The team began with 175 examples written by humans and used a modified Self-Instruct method with OpenAI's text-davinci-003 to expand them into 52,000 demonstrations. Fine-tuning LLaMA 7B on that set took about three hours on eight 80 GB Nvidia A100 GPUs.

How much did the project cost?

Producing the training data cost less than US$500, and the fine-tuning run was estimated at under US$100. That low total is a large part of why the project drew attention.

Can I use it commercially?

No. Stanford released the dataset and training code for research purposes. The code is Apache 2.0, but the dataset and weight diff are CC BY-NC 4.0, and the model inherits limits from LLaMA and from data generated through OpenAI's API.

What are its weaknesses?

The researchers reported that it can hallucinate, produce misinformation, use toxic language and repeat stereotypes, and they said it was not intended for general deployment. They also called their evaluation preliminary and noted flaws inherited from LLaMA and the training data.

Good reading and sources

Where this page comes from

The facts on this page come from the Wikipedia article "Stanford Alpaca", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryLLM rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.

It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

The weekly digest

Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.