Assisted fine adjustment
At DevDay last November, in collaboration with a dedicated group of OpenAI researchers, we announced a custom model program designed to train and optimize models for specific domains. Since then, we’ve met with dozens of customers, evaluated their custom model needs, and evolved the program to further maximize performance.
Today, we are officially announcing our fine-tuning assistance service as part of our Custom Models program. Supplemental Fine-Tuning is a collaboration with our technical team to leverage techniques beyond the fine-tuning API, such as additional hyperparameters and large-scale eclectic parameter-efficient fine-tuning (PEFT) methods. This is especially useful for organizations that require support for setting up efficient training data pipelines, evaluation systems, and bespoke parameters and methods to maximize model performance for use cases and tasks.
for example, SK Telecom, a telecommunications operator serving more than 30 million subscribers in South Korea, initially wanted to focus on customer service and customize its model to become an expert in the telecommunications field. They worked with his OpenAI to fine-tune his GPT-4 to improve its performance in communication-related conversations in Korean. Over several weeks, SKT and OpenAI significantly improved performance on telecom customer service tasks. Conversation summary quality improved by 35%, intent recognition accuracy improved by 33%, and satisfaction score increased from 3.6 to 4.5. 5) When comparing the fine-tuned model with his GPT-4.
custom trained model
In some cases, organizations need to train specialized models from scratch that understand their business, industry, and domain. A completely custom-trained model injects new knowledge from a specific domain by using new during- and post-training techniques to modify key steps in the model training process. Organizations that are successful with fully custom-trained models often have large amounts of proprietary data (millions of samples or billions of tokens) and are looking to use this data. We want to teach our models new knowledge and complex, unique behaviors for very specific use cases.
for example, harvey, an AI-native legal tool for lawyers, has partnered with OpenAI to create a custom-trained large-scale language model for case law. Although the foundation model was good for reasoning, it lacked extensive knowledge of case history and other knowledge needed for legal work. After testing his engineering prompts, RAGs, and tweaks, Harvey worked with his team to add the necessary depth of context to the model. This equates to his 10 billion tokens worth of data. Our team has modified every step of the model’s training process, from during domain-specific training to customizing the post-training process to incorporating attorney expert feedback. The resulting model increased his factual responses by 83%, and the lawyer said he preferred the customized model’s output to GPT-4 97% of the time.