2024-04-05 13:01:12
OpenAI has announced the implementation of new customization options for the application programming interface (API) to adjust its Artificial Intelligence (AI) models, such as the possibility of connecting with third-party platforms to share information, in addition to expanding its program of custom models.
The AI company made the fine tuning API of its GPT-3.5 Turbo language model available to developers in August of last year. Since that time, OpenAI has noted that “thousands of organizations have trained hundreds of thousands of models using its API.”
This is an API that allows users to customize a model uniquely for their application. Thus, with the tuning API, the model is retrained by adding additional data from a personal database to update its parameters.
In this way, the API allows you to get more out of AI models, managing to offer more precise and higher quality results, as well as the ability to manage lower latency requests and, therefore, save on the use of ‘tokens’, which translates into a cheaper cost.
Now, the technology company led by Sam Altman has announced the launch of more features for developers, which will provide more control over API settings and, thereby, improve the accuracy of AI and reduce costs.
In this sense, as detailed in a statement on its website, one of these new functions is the option to integrate with third-party platforms, allowing developers to share detailed adjustment data with other providers. Specifically, this option will begin to be available with the Weights and Biases model development platform.
It will also allow you to create control points according to the model’s training periods. With this, developers will be able to save adjusted models with the API so as not to have to retrain them later, especially in cases of overfitting.
OpenAI has launched a new parallel user interface, which will allow the performance and quality of the model in question to be compared. And it has introduced the ability to calculate end-to-end validation metrics, for example metrics such as response accuracy across the entire data set, providing better insight into model quality.
Finally, the company has added improvements to the dashboard, such as the ability to configure hyperparameters, view more detailed training metrics, and rerun jobs from previous configurations.
EXPANSION OF THE CUSTOM MODELS PROGRAM
On the other hand, OpenAI has detailed that another option to increase AI model performance is to build a custom trained model. To this end, they have announced that they are expanding the personalized models program, introducing more ways to work with the company’s AI researchers and experts.
As indicated by OpenAI, they have implemented an assisted adjustment option as part of the program, with which organizations will be able to collaborate with the company’s technical teams to take advantage of “techniques beyond the adjustment API.”
Specifically, the Custom Models program gives select organizations the opportunity to work with a dedicated group of technology researchers to train GPT-4 models customized for their specific domain. Thus, this program is applicable to organizations with “extremely large” data sets. That is, at a minimum, they must manage billions of ‘tokens’.
All in all, this is a useful feature for organizations that require support in configuring issues such as evaluation systems and parameters, or efficient custom methods to maximize model performance for their use case or task.
Likewise, this program also includes those organizations that need to train from scratch a specifically designed model so that, for example, the AI understands specific points such as their type of business, industry or sector. “Fully custom-trained models provide new knowledge of a specific domain,” the technology company stated.
1712327899
#OpenAI #Adds #Customization #Options #Fitting #API #Expands #Custom #Model #Program