The fine-tuning process involves creating a fine-tuning job and using the fine-tuned model, with a fine-tuning UI to be launched soon. Safety is a priority, with fine-tuning training data passed through a Moderation API and a GPT-4 powered moderation system. The cost of fine-tuning is split into training and usage costs. OpenAI has also announced the availability of `babbage-002` and `davinci-002` as replacements for the original GPT-3 base models, which can be fine-tuned using a new API endpoint. The old endpoint will be turned off on January 4th, 2024.
Key takeaways:
- Fine-tuning for GPT-3.5 Turbo is now available, allowing developers to customize models for better performance in their specific use cases. Fine-tuning for GPT-4 is expected to be available this fall.
- Use cases for fine-tuning include improved steerability, reliable output formatting, and custom tone. Fine-tuning also allows for shorter prompts and can handle 4k tokens.
- Fine-tuning costs are divided into initial training cost and usage cost, with specific rates per 1K tokens for training, usage input, and usage output.
- New GPT-3 models, `babbage-002` and `davinci-002`, are now available as replacements for the original GPT-3 base models. The old `/v1/fine-tunes` endpoint will be deprecated and turned off on January 4th, 2024.