Assisted fine tuning

At DevDay last November, we announced a custom model program designed to train and optimize domain-specific models, in collaboration with a dedicated group of OpenAI researchers. Since then, we have met with dozens of customers to assess their custom model needs and develop our program to further increase performance.

Today we are officially announcing our assisted fine-tuning offering as part of the Custom Model program. Assisted Fine Tuning is a collaborative effort with our technical teams to leverage techniques outside of the Fine Tuning API, such as additional hyperparameters and various Parameter Efficiency Fine Tuning (PEFT) methods on a larger scale. It is particularly useful for organizations that need support in setting up effective training data channels, evaluation systems, and custom parameters and methods to maximize model performance for their use case or task.

For example, SK Telecom, a telecommunications operator serving more than 30 million subscribers in South Korea, wanted to adapt the model to be a specialist in the telecommunications domain with an initial focus on customer service. They collaborated with OpenAI to fine-tune GPT-4 to improve its performance in Korean-language telecommunications-related conversations. Over several weeks, SKT and OpenAI led to significant performance improvements in telecom customer service tasks – a 35% increase in conversation summary quality, a 33% increase in intent recognition accuracy, and an increase in satisfaction scores from 3.6 to 4.5 (from 5) when comparing the fine-tuned model with GPT-4.

Custom model

In some cases, organizations must train from scratch a purpose-built model that understands their business, industry or domain. Fully customized models infuse new domain knowledge by modifying key steps of the model training process using new mid-training and post-training techniques. Organizations that see success with a fully custom model often have large amounts of proprietary data—millions of instances or billions of tokens—that they want to use to teach the model new knowledge or complex, unique behaviors for very specific use cases.

For example, HarveyAn AI native legal tool for lawyers, in partnership with OpenAI for create a custom large language model for case law. Although the foundation models were strong in reasoning, they lacked extensive knowledge of the legal history of the case and other knowledge required for legal work. After testing rapid engineering, RAG and fine-tuning, Harvey worked with our team to add the depth of context needed to the model – the equivalent of 10 billion tokens worth of data. Our team modified each step of the training process model, from domain-specific intermediate training to post-training process adjustments and incorporating expert attorney feedback. The resulting model achieved an 83% increase in factual responses, and lawyers preferred the results of the adjusted model over the GPT-4 97% of the time.