Photo credit: venturebeat.com
OpenAI has recently revealed its latest initiative that permits external software developers to fine-tune custom iterations of its advanced large multimodal model (LMM), GPT-4o. This new feature aims to adapt the model’s capabilities to better meet specific application or organizational needs.
Fine-tuning provides an opportunity for developers to refine aspects such as tone, adherence to directives, and enhancement of accuracy in specialized tasks. Even minimal datasets can yield substantial improvements through this process.
On the newly introduced fine-tuning dashboard, developers can easily create custom models by selecting “create” and choosing gpt-4o-2024-08-06 from the base model options. Interested developers can find further instructions on OpenAI’s fine-tuning dashboard.
This announcement follows another recent update, wherein OpenAI enabled fine-tuning capabilities for the more compact and cost-effective variant, GPT-4o mini, providing a lighter alternative although at reduced power compared to the full-scale GPT-4o.
OpenAI’s technical team, comprising John Allard and Steven Heidel, highlighted the transformative potential of fine-tuning across a diverse range of fields in a company blog post. They noted, “This is just the beginning; we’re committed to enhancing our custom model options for developers.”
Free Tokens Available Until September 23
OpenAI points out that effective results in fine-tuning can be achieved using only a handful of examples in training datasets.
To encourage adoption of this feature, the company is granting access to 1 million free tokens per day for fine-tuning GPT-4o for any external organization up until September 23, 2024.
Tokens represent the numerical values associated with various text elements, encapsulating the learned concepts within an LLM or LMM. They serve as a performance measure for how much data the model processes (input) or generates (output). Developers looking to fine-tune GPT-4o must translate relevant organizational data into tokens compatible with the model, a process facilitated by OpenAI’s fine-tuning tools.
Typically, the cost to fine-tune GPT-4o sits at $25 per million tokens, with running the refined model costing $3.75 for every million input tokens and $15 per million output tokens.
In contrast, free training tokens are available daily for the GPT-4o mini model, amounting to 2 million until September 23. This initiative ensures widespread access for developers operating within paid usage tiers.
This move to provide complimentary tokens arrives amid aggressive pricing competition from other major players like Google and Anthropic, in addition to open-source alternatives such as the newly launched Hermes 3 from Nous Research, a variant of Meta’s Llama 3.1.
Nonetheless, opting for OpenAI’s models alleviates concerns regarding local hosting for model inference or training since developers can utilize OpenAI’s infrastructure or connect their own preferred servers to OpenAI’s API.
Demonstrating Fine-Tuning Success
The introduction of GPT-4o fine-tuning is the result of extensive collaboration with select partners, showcasing the capabilities of customized models across different sectors.
Notably, Cosine, a firm specializing in AI software engineering, has successfully employed fine-tuning, achieving a state-of-the-art result of 43.8% on the SWE-bench benchmark through its autonomous AI engineer agent, Genie. This marks the highest performance recorded for publicly acknowledged AI models to date.
Similarly, Distyl, an AI solutions provider for Fortune 500 companies, reported outstanding results with its fine-tuned GPT-4o, claiming the top spot on the BIRD-SQL benchmark with an impressive execution accuracy of 71.83%. The model showed particular proficiency in tasks such as query reformulation, intent classification, chain-of-thought reasoning, and SQL generation self-correction.
Prioritizing Safety and Data Privacy
OpenAI remains steadfast in its commitment to safety and data privacy, even as it broadens the customization options for developers.
The fine-tuned models ensure complete control over business data, guaranteeing that input and output data are not utilized in training other models.
To reinforce these principles, the company has established multiple safety measures, including automated evaluations and active monitoring, to monitor compliance with OpenAI’s usage policies.
However, research indicates that fine-tuning can sometimes lead models to stray from established safety protocols and may compromise their overall performance. The decision of whether the benefits outweigh the risks rests with individual organizations; however, OpenAI is certainly advocating for the value of fine-tuning.
When OpenAI introduced fine-tuning tools for developers in April—such as epoch-based checkpoint creation—it suggested that “in the future, most organizations will develop personalized models tailored to their industry, business, or application.”
The recent release of GPT-4o fine-tuning features reinforces OpenAI’s ongoing mission to realize a future where every organization can possess its own specialized AI model.
Source
venturebeat.com