New features will help developers have more control over ChatGPT fine-tuning and provide new ways to build custom models with OpenAI.

OpenAI’s fine-tuning API for GPT-3.5, launched in August 2023, has recently received an update with the new features.

Fine-tuning API is dedicated to training large language models (LLMs) with more examples than can fit in a single prompt. Fine-tuning helps LLMs deeply understand content and leverage existing knowledge for a specific task.

One of the most important improvements is the epoch-based checkpoint creation. It enables developers to automatically produce one full fine-tuned model checkpoint during each training epoch. Therefore, the need for subsequent retraining decreases, especially in the cases of overfitting.

Another novel feature is the side-by-side Playground UI. It allows for comparing model quality and performance, introducing human evaluation of the outputs of multiple models or fine-tuning snapshots against a single prompt.

From now on, the fine-tuning API will also have support for integrations with third-party platforms (starting with the Weights and Biases AI developer platform).

The updated version will also provide more insights with comprehensive validation metrics. The ability to compute metrics like loss and accuracy over the entire validation dataset instead of a sampled batch shall give better awareness and assessment of model quality.

The new design of the API dashboard enables developers to configure hyperparameters, view more detailed training metrics, and rerun jobs from previous configurations.

Besides the API update, OpenAI also announced the expansion of its Custom Model program designed to train and optimize models for a specific domain. Now the program will include assisted fine-tuning which means the use of techniques beyond the fine-tuning API, such as additional hyperparameters and various parameter-efficient fine-tuning methods at a larger scale.

Therefore, the partner companies may now include new, domain-specific knowledge into the model to train a purpose-built model that understands their business, industry, or domain from scratch.

As we have previously reported, OpenAI currently witnesses an increase in consumer demand for the corporate version of its ChatGPT chatbot. During a conversation with media representatives, OpenAI Chief Operating Officer Brad Lightcap stated that nowadays more than 600,000 people are ChatGPT Enterprise users.