Book a call

News · December 14, 2021

OpenAI lets every API customer fine-tune GPT-3 on their own data

// ai.news :: 2021

Developers can now train a custom version of GPT-3 with their own examples using a single command. For many business tasks, a tuned model may be more accurate and cheaper to run than long prompts.

What happened

  • OpenAI announced on December 14, 2021 that all API customers can fine-tune GPT-3 on their own data, starting with one command in its command-line tool, and that the custom model becomes available in the API once trained.
  • OpenAI said benefits can appear with fewer than 100 examples and that, in its grade-school math tests, fine-tuning improved accuracy by two to four times compared with prompt design alone.
  • According to OpenAI, Viable raised the accuracy of its customer-feedback summaries from 66% to 90%, and Keeper Tax improved accuracy from 85% to 93% by adding training examples weekly.
  • VentureBeat reported that OpenAI charges a premium for fine-tuned models but that they often need shorter prompts, which can reduce token costs.

Why it matters

Fine-tuning turns GPT-3 from a general tool into something closer to a component trained on your own domain. Support answers, classification of tickets or reviews and structured extraction are good candidates, because they have clear right and wrong answers to train on.

The real work is the data: a few hundred clean, reviewed examples, a test set kept aside, and a habit of adding new examples as the product changes. Compare the tuned model against a good prompt on cost and accuracy before committing, and remember that sending data for training needs to fit your privacy commitments to users.

Sources

  • OpenAI: Customizing GPT-3 for your application
  • VentureBeat: OpenAI begins allowing customers to fine-tune GPT-3

This post is CMM’s own summary and commentary on publicly reported news. Facts are attributed to the sources linked above; please read them for full details. CMM is not affiliated with these publishers or with the companies mentioned, and trademarks belong to their owners. Prepared with AI assistance. Corrections or requests: [email protected]. See our editorial policy.

Want help with this?

Tell us what you need, and we will suggest the right team model on a short call.

Talk to us