OpenAI Introduces Fine-Tuning and API Updates for GPT-3.5 Turbo

OpenAI has unveiled fine-tuning capabilities for GPT-3.5 Turbo, a step that empowers developers to customize models according to their specific use cases. The fine-tuning for GPT-3.5 Turbo is available now, with an extension to GPT-4 slated for this fall. This development is pivotal as it not only allows customization of models but also enhances performance at scale. Early indications reveal that a fine-tuned GPT-3.5 Turbo can rival or even surpass the base GPT-4 model in certain specialized tasks.

Fine-Tuning Use Cases

With the release of GPT-3.5 Turbo, there was a growing demand from developers and businesses for model customization to create distinct user experiences. This update addresses that demand by enabling supervised fine-tuning, which significantly improves model performance across various domains:

Improved Steerability: Fine-tuning helps in better instruction following, for instance, ensuring the model responds in a specified language or adheres to a particular response format.

Reliable Output Formatting: It enhances the model's capability to format responses consistently, which is vital for tasks like code completion or composing API calls.

Custom Tone: Businesses can fine-tune the model to align with their brand's voice, ensuring consistency in tone across interactions.

Moreover, fine-tuning allows businesses to shorten their prompts while maintaining performance levels, and it supports handling up to 4k tokens, which is double the capacity of previous models. The outcome is a reduction in prompt size by up to 90%, accelerating API calls and reducing costs.

Fine-Tuning Process

The fine-tuning process is straightforward:

Prepare Your Data: Get your data ready for the fine-tuning process.

Upload Files: Upload the necessary files for fine-tuning.

Create a Fine-Tuning Job: Initiate the fine-tuning job.

Use a Fine-Tuned Model: Once the process is complete, the fine-tuned model is ready for production use.

An upcoming fine-tuning UI will further simplify the process by providing easy access to information about ongoing fine-tuning jobs and completed model snapshots.

Safety Measures

OpenAI ensures the safety of the fine-tuning process by employing a Moderation API and a GPT-4 powered moderation system to identify unsafe training data, thereby preserving the default model's safety features.

Pricing

The pricing for fine-tuning is bifurcated into training costs and usage costs, making it transparent and easy to understand for developers. For example, a fine-tuning job with a training file of 100,000 tokens trained for 3 epochs would cost $2.40.

Updated GPT-3 Models

OpenAI has also introduced updated versions of original GPT-3 base models, babbage-002 and davinci-002, which can be accessed via the Completions API. They can be fine-tuned using the new API endpoint /v1/fine_tuning/jobs, offering more extensibility to support the future evolution of the fine-tuning API.

This suite of updates marks a significant milestone in making GPT-3.5 Turbo more adaptable and useful to developers, paving the way for more personalized and efficient AI applications.