OpenAI has released fine-tuning capabilities for GPT-3.5 Turbo, allowing developers to tailor the model to their use cases. Highlights of this update include:
Fine-tuning possibilities: GPT-3.5 Turbo’s fine-tuning allows developers to better adapt the model to their use cases. This can improve the format, language, and tone of the output, and reduce prompt size by up to 90%.
Safety Guarantee: To maintain the safety function of the base model during the fine-tuning process, we detect risky training data that conflicts with safety criteria through OpenAI’s Moderation API and a GPT-4-based censorship system.
Pricing Structure: The cost of fine-tuning is divided into the initial training cost and the usage cost. For example, the estimated cost of a gpt-3.5-turbo fine-tuning job trained for 3 epochs with a training file of 100,000 tokens is $2.40.
New models and endpoints: New models babbage-002 and davinci-002 are now available, along with a new API endpoint /v1/fine_tuning/jobs. The old endpoint will be deprecated on January 4, 2024.
Future support: Function calling and fine-tuning support for gpt-3.5-turbo-16k will be added in the fall, with a fine-tuning UI coming soon.
With these updates, OpenAI is giving developers the tools to leverage GPT-3.5 Turbo more effectively, demonstrating its ability to reach or surpass baseline GPT-4 performance on certain tasks .
In the future, this kind of LLM utilization will become easier and easier. GPT-4 is also scheduled to release its fine-tuning function in the fall. Users will have to think about how to use this intelligently, and those developing LLMs will likely face a crossroads of whether to create increasingly lightweight models or models specialized for specific areas. It feels like the market standard for Foundation Models that cover all aspects is getting higher and higher.
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. Would you like to be notified when new articles are posted? 🔔 Yes, that means subscribe. haebom@kakao.com
