GPT-3.5 Turbo fine-tuning features

Customers of OpenAI now have the option to introduce custom datasets to GPT-3.5 Turbo, the streamlined iteration of the GPT-3.5 model. This capability facilitates the enhancement of the AI model's textual generation reliability while incorporating distinct behavioral patterns.

OpenAI asserts that refined iterations of GPT-3.5 exhibit the potential to equal or surpass the foundational competencies of GPT-4, the company's primary model, in "specific specialized tasks."

Since the introduction of GPT-3.5 Turbo, there has been an increasing desire among developers and enterprises to incorporate personalized features into the model in order to curate distinctive and unique engagements for their user base. A blog post published today expounded on this development, explaining that "This enhancement furnishes developers with the ability to fine-tune models for heightened performance tailored to their precise use cases, and subsequently leverage these customized models at a notable magnitude."

By means of fine-tuning, enterprises harnessing GPT-3.5 Turbo via OpenAI’s API can refine the model's compliance with directives, an example being the establishment of an unwavering preference for responding in a designated language. Alternatively, organizations can enhance the model's capacity for maintaining consistent response formats (e.g., when concluding code segments), while also refining the model's output nuances, including its demeanor and tone, to seamlessly align with a brand's identity or distinct voice.

Additionally, the implementation of fine-tuning bestows upon OpenAI's patrons the ability to condense their textual prompts, leading to accelerated API invocation and cost efficiency. OpenAI's blog entry asserts that "early adopters have successfully reduced prompt dimensions by as much as 90% through the strategic integration of fine-tuned instructions within the model itself."

The process of fine-tuning presently involves data preparation, file upload, and the initiation of a fine-tuning task via OpenAI's API. All data intended for fine-tuning undergoes assessment through a 'moderation' API and a GPT-4-driven moderation mechanism to ensure alignment with OpenAI's safety protocols, according to the company's statement. OpenAI has outlined its intentions to introduce a dedicated fine-tuning user interface (UI) in the future, complete with a comprehensive dashboard to facilitate real-time monitoring of ongoing fine-tuning initiatives.

The cost structure for fine-tuning is outlined below:

• Training cost amounts to $0.008 per 1,000 tokens.
• For usage input, the fee is $0.012 per 1,000 tokens
• For usage output, the cost is $0.016 per 1,000 tokens

"Tokens" signifies unprocessed text fragments, such as "fan," "tas," and "tic," which together form the word "fantastic." A fine-tuning task for GPT-3.5 Turbo involving a training file of 100,000 tokens, approximately equating to 75,000 words, is estimated to incur a cost of around $2.40, as specified by OpenAI.

In recent developments, OpenAI has introduced two revised GPT-3 base models, namely babbage-002 and davinci-002. These models are now equipped with the capability of undergoing fine-tuning and come with enhanced features such as pagination and extended adaptability. As previously communicated, OpenAI is set to retire the initial GPT-3 base models on January 4, 2024.

OpenAI has communicated that the introduction of fine-tuning support for GPT-4 is on the horizon. Distinguishing itself from GPT-3.5, GPT-4 possesses the ability to comprehend images alongside textual content. The anticipated rollout of this feature is projected for later in the fall, although precise details regarding its implementation remain forthcoming.

Source