Sign In

Now you can tailor GPT-3.5 to your needs more easily than ever.

Haebom
OpenAI has launched fine-tuning for GPT-3.5 Turbo, enabling developers to customize the model for their own use cases. The main highlights of this update are as follows.
Fine-tuning possibilities: Fine-tuning for GPT-3.5 Turbo lets developers adapt the model even more precisely to their needs. This allows improvements to the output's format, language, and tone, and can reduce prompt size by up to 90%.
Ensuring safety: During fine-tuning, OpenAI’s Moderation API and a GPT-4-based filtering system are used to detect risky training data that may conflict with safety standards, maintaining the safety features of the base model.
Pricing structure: The cost of fine-tuning is separated into the initial training cost and usage cost. For instance, training gpt-3.5-turbo with 100,000-token files over three epochs is estimated to cost $2.40.
New models and endpoints: The babbage-002 and davinci-002 models are now newly released, with a new API endpoint /v1/fine_tuning/jobs provided. The previous endpoint will be discontinued on January 4, 2024.
Future support: Support for function calling and fine-tuning for gpt-3.5-turbo-16k is scheduled for the fall, and a fine-tuning UI will be launched soon.
With these updates, OpenAI now offers developers the tools to make even more effective use of GPT-3.5 Turbo, achieving or even exceeding the base performance of GPT-4 on certain tasks,demonstrating this capability.
Moving forward, it will get easier and easier to utilize LLMs like this. Fine-tuning for GPT-4 is reportedly planned for this fall as well. Users will need to figure out how to use these tools wisely, while developers of LLMs may have to decide whether to build more lightweight models or ones specialized for specific domains. It seems like the expectations for all-encompassing Foundation Models are rising across the market.
Subscribe to 'haebom'
📚 Welcome to Haebom's archives.
---
I post articles related to IT 💻, economy 💰, and humanities 🎭.
If you are curious about my thoughts, perspectives or interests, please subscribe.
haebom@kakao.com
Subscribe