Assisted fine adjustment
At DevDay last November, we announced a custom model program designed to train and optimize models for a specific domain, in partnership with a dedicated group of OpenAI researchers. Since then, we have met with dozens of customers to evaluate the needs of their custom models and have developed our program to further maximize performance.
Today we formally announced our assisted fitting offering as part of the Custom Model program. Assisted tuning is a collaborative effort with our technical teams to leverage techniques beyond the tuning API, such as additional hyperparameters and various larger scale parameter efficient tuning (PEFT) methods. It is particularly useful for organizations that need support in setting up training data pipelines, evaluation systems, and efficient custom parameters and methods to maximize model performance for their use case or task.
For example, SK Telecom, a telecom operator serving over 30 million subscribers in South Korea, wanted to customize a model to be an expert in the telecom space with an initial focus on customer service. They worked with OpenAI to tune GPT-4 and improve its performance in Korean-language telecom-related conversations. Over the course of several weeks, SKT and OpenAI drove significant performance improvement in telecom customer service tasks: a 35% increase in conversation summary quality, a 33% increase in voice recognition accuracy, intentions and an increase in satisfaction scores from 3.6 to 4.5 (out of 5) when comparing the adjusted model with GPT-4.
Custom trained model
In some cases, organizations need to train from scratch a specifically designed model that understands their business, industry, or domain. Fully custom-trained models provide new insights into a specific domain by modifying key steps in the model training process using novel techniques mid- and post-training. Organizations that see success with a fully customized model often have large amounts of proprietary data (millions of examples or billions of tokens) that they want to use to teach the model complex and unique new insights or behaviors for very specific use cases. .
For example, ai/” rel=”noopener noreferrer” target=”_blank”>Harvey, a native ai legal tool for lawyers, partnered with OpenAI to create a custom large language model for case law. While the basic models were sound in reasoning, they lacked the broad knowledge of legal case history and other knowledge necessary for legal work. After testing rapid engineering, RAG, and tuning, Harvey worked with our team to add the necessary depth of context to the model: the equivalent of 10 billion tokens worth of data. Our team modified every step of the model training process, from intermediate training in a specific domain to customizing post-training processes and incorporating feedback from expert attorneys. The resulting model achieved an 83% increase in factual responses, and attorneys preferred the results from the custom model 97% of the time over GPT-4.
