OpenAI Announces New Fine-Tuning Features, Expands Custom Model Program

OpenAI Announces New Fine-Tuning Features, Expands Custom Model Program

OpenAI is ramping up efforts to help enterprises develop tailored AI models with the expansion of its Custom Model program and the introduction of assisted fine-tuning. This move comes as the company is experiencing rapid growth in the adoption of its enterprise-focused ChatGPT offering, with Chief Operating Officer Brad Lightcap revealing that the number of users signed up for ChatGPT Enterprise has quadrupled since January, now exceeding 600,000.

OpenAI Sees Surging Demand for Enterprise as ChatGPT Adoption Soars
ChatGPT Enterprise now boasts over 600,000 users, a remarkable increase from the 150,000 users in January.

The Custom Model program, initially launched last year at OpenAI's inaugural DevDay developer conference, has already attracted dozens of customers seeking to work directly with OpenAI researchers to train and optimize models for their unique requirements. The program's success has prompted OpenAI to expand its offerings to better serve the needs of its enterprise clients.

OpenAI launched its self-serve fine-tuning API for GPT-3.5 in August 2023, which has since been used by thousands of organizations to train hundreds of thousands of models. The new features for the fine-tuning API include epoch-based checkpoint creation, a new side-by-side Playground UI for comparing model quality and performance, support for integrations with third-party platforms (starting with Weights and Biases this week), comprehensive validation metrics, the ability to configure available hyperparameters from the dashboard, and more.

One of the key additions to the program is assisted fine-tuning, a collaborative effort between OpenAI's technical teams and customers to employ advanced techniques beyond the standard fine-tuning API. This includes the use of additional hyperparameters and various parameter-efficient fine-tuning (PEFT) methods at a larger scale. Assisted fine-tuning is particularly beneficial for organizations that require support in setting up efficient training data pipelines, evaluation systems, and bespoke parameters to maximize model performance for their specific use case or task.

SK Telecom, a telecommunications operator in South Korea, has already seen significant improvements in their customer service AI through assisted fine-tuning. By working with OpenAI to fine-tune GPT-4 for telecom-related conversations in Korean, SK Telecom achieved a 35% increase in conversation summarization quality, a 33% increase in intent recognition accuracy, and a notable increase in satisfaction scores compared to the base GPT-4 model.

For organizations with highly specific use cases and large quantities of proprietary data, OpenAI offers the option of fully custom-trained models. These models are built from scratch, incorporating domain-specific knowledge through novel mid-training and post-training techniques.

Harvey, an AI-native legal tool for attorneys, partnered with OpenAI to create a custom-trained large language model for case law. By modifying every step of the model training process and incorporating the equivalent of 10 billion tokens worth of legal data, the resulting model achieved an 83% increase in factual responses, with attorneys preferring its outputs 97% of the time over GPT-4.

OpenAI believes that the future of AI lies in the development of customized models tailored to specific industries, businesses, and use cases. With a range of techniques available, from the self-serve fine-tuning API to the Custom Model program, organizations of all sizes can now create personalized AI models to drive more meaningful and specific impact from their AI implementations.

As the demand for tailored AI solutions continues to grow, OpenAI's expanded Custom Model program and the introduction of assisted fine-tuning are set to play a crucial role in helping enterprises harness the full potential of AI technology.

Let’s stay in touch. Get the latest AI news from Maginative in your inbox.