Unlock New AI Potential: Enhanced Fine-Tuning API & Custom Models Program Expansion

Artificial Intelligence (AI) continues to revolutionize industries by providing innovative solutions to complex problems. At OpenAI, we are committed to pushing the boundaries of what AI can achieve. Today, we are excited to announce significant improvements to our fine-tuning API and the expansion of our custom models program. These enhancements are designed to give developers more control over their AI models and provide new ways to build custom models tailored to specific domains.

New Fine-Tuning API Features

Since the launch of our self-serve fine-tuning API for GPT-3.5 in August 2023, thousands of organizations have leveraged it to train hundreds of thousands of models. Fine-tuning allows models to deeply understand content and augment their existing knowledge and capabilities for specific tasks. This can lead to higher quality results while reducing costs and latency.

Key Enhancements

Epoch-based Checkpoint Creation: This feature automatically produces a full fine-tuned model checkpoint during each training epoch. This reduces the need for subsequent retraining, especially in cases of overfitting, and ensures that you always have a reliable model version to fall back on.
Comparative Playground: Our new side-by-side Playground UI allows developers to compare model quality and performance. This feature enables human evaluation of the outputs of multiple models or fine-tune snapshots against a single prompt, making it easier to choose the best-performing model.
Third-party Integration: We now support integrations with third-party platforms, starting with Weights and Biases. This allows developers to share detailed fine-tuning data with the rest of their stack, facilitating better collaboration and analysis.
Comprehensive Validation Metrics: Developers can now compute metrics like loss and accuracy over the entire validation dataset instead of a sampled batch. This provides better insight into model quality and helps in making more informed decisions.
Hyperparameter Configuration: You can now configure available hyperparameters directly from the Dashboard, rather than only through the API or SDK. This makes it easier to experiment with different configurations and optimize your model.
Fine-Tuning Dashboard Improvements: The Dashboard now includes the ability to configure hyperparameters, view more detailed training metrics, and rerun jobs from previous configurations. These improvements make it easier to manage and optimize your fine-tuning jobs.

Real-World Applications

Fine-tuning has a wide range of applications. For example, Indeed, a global job matching and hiring platform, used our fine-tuning API to generate personalized job recommendations. By fine-tuning GPT-3.5 Turbo, Indeed was able to improve the quality and accuracy of their recommendations, reduce the number of tokens in prompts by 80%, and scale from less than one million messages to job seekers per month to roughly 20 million.

Expanding Our Custom Models Program

In addition to enhancing our fine-tuning API, we are also expanding our custom models program. This program is designed to help organizations build models that are optimized for specific domains, in partnership with our team of AI experts and researchers.

Assisted Fine-Tuning

Assisted fine-tuning is a collaborative effort with our technical teams to leverage techniques beyond the fine-tuning API. This includes additional hyperparameters and various parameter-efficient fine-tuning (PEFT) methods at a larger scale. It is particularly helpful for organizations that need support setting up efficient training data pipelines, evaluation systems, and bespoke parameters and methods to maximize model performance for their use case.

For example, SK Telecom, a telecommunications operator serving over 30 million subscribers in South Korea, worked with OpenAI to fine-tune GPT-4 for telecom-related conversations in the Korean language. This collaboration led to a 35% increase in conversation summarization quality, a 33% increase in intent recognition accuracy, and an increase in satisfaction scores from 3.6 to 4.5 (out of 5).

Custom-Trained Models

In some cases, organizations need to train a purpose-built model from scratch that understands their business, industry, or domain. Fully custom-trained models imbue new knowledge from a specific domain by modifying key steps of the model training process using novel mid-training and post-training techniques.

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

What’s Next for Model Customization

We believe that in the future, the vast majority of organizations will develop customized models that are personalized to their industry, business, or use case. With a variety of techniques available to build a custom model, organizations of all sizes can develop personalized models to realize more meaningful, specific impact from their AI implementations.

Steps to Success

Clearly Scope the Use Case: Define the specific problem you want to solve and the desired outcomes. This will help in choosing the right techniques and data for your model.
Design and Implement Evaluation Systems: Set up robust evaluation systems to measure the performance of your model. This includes metrics like accuracy, loss, and user satisfaction scores.
Choose the Right Techniques: Depending on your use case, you may need to use fine-tuning, retrieval-augmented generation (RAG), or custom-trained models. Each technique has its strengths and can be used to achieve different goals.
Iterate Over Time: AI models often require multiple iterations to reach optimal performance. Be prepared to refine your model based on feedback and new data.

Getting Started

With OpenAI, most organizations can see meaningful results quickly with the self-serve fine-tuning API. For organizations that need to more deeply fine-tune their models or imbue new, domain-specific knowledge into the model, our Custom Model programs can help.

Visit our fine-tuning API documentation to start fine-tuning our models. For more information on how we can help customize models for your use case, reach out to us.

Conclusion

The enhancements to our fine-tuning API and the expansion of our custom models program represent significant steps forward in making AI more accessible and effective for a wide range of applications. Whether you are looking to fine-tune an existing model or build a custom-trained model from scratch, OpenAI provides the tools and expertise to help you achieve your goals.

By leveraging these new features and programs, developers and organizations can unlock new AI potential, improve model performance, and create more personalized and impactful AI solutions. We are excited to see what you will build with these new capabilities and look forward to continuing to support your AI journey.

For more information, visit our website or contact us directly.