OpenAI has launched tools to customise GPT-3. Developers can fine-tune GPT-3 on their data and create a customised version tailored to their application. Such customising will make GPT-3 reliable for wider use cases, and running the model becomes cheaper and faster.
Just running a single command in the OpenAI command-line tool with a file will start training the custom model and make it available immediately in API. Developers can use an existing dataset of virtually any size and shape or incrementally add data based on user feedback. According to the company, customisation can help in increasing correct outputs from 83% to 95%.
According to research by OpenAI, in less than 100 examples, the benefits of fine-tuning GPT-3 are visible, and the performance continues to improve as more data is added. A customised GPT-3 model with Grade School Math problems dataset improved accuracy by 2 to 4x over what’s possible with prompt design. Developers can improve the reliability of output and offer more consistent results by customising GPT-3.
As custom versions of GPT-3 are tailored to specific required applications, the prompt can be much short, with improved latency and reduced costs. Be it summarization, text generation, classification, or any other natural language task that GPT-3 is capable of performing; a customised GPT-3 will improve performance.
Some of the apps powered by customised versions of GPT-3 include Keeper Tax, Viable, Sana Labs, and Elicit. Keeper Tax adds 500 new training examples every week to fine-tune their model, which has led to an increase in accuracy from 85% to 93%. Viable shows better results, where with a customised version of GPT-3, accuracy in summarising customer feedback has improved from 66% to 90%.
OpenAI’s all API customers have access to customised GPT-3.