The developer of the well-known AI chatbot ChatGPT, OpenAI, has announced the availability of fine-tuning features for its GPT-3.5 Turbo model. Developers will be able to run the current model at scale and modify it according to their use cases, thanks to the new feature.
According to OpenAI, the fine tuned GPT-3.5 Turbo can even surpass the performance of GPT-4. In preliminary tests, a refined version of the model was able to match, or even outperform, base GPT-4-level capabilities on certain narrow tasks, according to OpenAI.
In order to increase a language model’s performance on a certain job, it can be fine-tuned by training it on a particular dataset. With OpenAI’s most recent development, companies can now fine-tune GPT 3.5 Turbo using their own datasets.
Read More: UK to Invest £100m in AI Chips Production Amid Global Competition
Businesses using the GPT-3.5 Turbo through the OpenAI API can fine-tune the model to make it more compliant, for example, by programming it to always reply in a specific language. Alternatively, they can enhance the model’s capacity to format responses consistently, for example, when completing snippets of code, and fine-tune the demeanor of the model’s output, such as its tone, to make it better fit a brand or voice.
Additionally, fine-tuning allows OpenAI users to reduce the length of their text prompts, accelerating API calls and lowering costs. According to OpenAI’s blog post, early testers have reduced prompt size by up to 90% by fine-tuning instructions into the model itself.
In order to fine-tune, one must currently prepare the data, upload the required files, and create a fine-tuning job using OpenAI’s API. The company claims that all fine-tuning data must go through a “moderation” API and a moderation system that is driven by GPT-4 to determine whether it violates OpenAI’s safety parameters.
But in the future, OpenAI intends to release a fine-tuning UI with a dashboard for monitoring the progress of ongoing fine-tuning workloads. OpenAI stated that fine-tuning support for GPT-4, which, in contrast to GPT-3.5, can recognise visuals in addition to text, will be available later this autumn.
