The right language model for your own application: Mozilla.ai wants to help
The project initially supports the selection of suitable LLMs and, in the foreseeable future, integration into the project and fine-tuning.
Mozilla.ai has introduced Lumigator, a framework designed to help integrate large language models (LLMs) into your own projects.
As a first step, Lumigator helps developers to select the right LLM. In the medium term, the plan is to develop the framework into an open source project for "ethical and transparent AI development", as stated in the blog post with the announcement.
The right model and the agony of choice
The declared aim of Lumigator is to make the selection of language models transparent, powerful and efficient. The first iteration offers a framework for evaluating LLMs based on task-specific metrics, enabling users to select the appropriate model for integration into their own applications.
In the blog post, the start-up Mozilla.ai, which was founded in early 2023 and operates independently of the Mozilla Foundation and Mozilla Corp, speaks of a reliable compass for the AI journey and cites the Artificial Intelligence Index Report 2024 of the Stanford Institute for Human-Centered Artificial Intelligence (HAI). Among other things, the study concludes that there is still a lack of tools for evaluating the models and that testing across several benchmarks complicates the comparison.
Lumigator wants to close this gap and prevent development teams from selecting the model based solely on the current hype and therefore often not using the LLM that is best suited to their specific requirements.
Big plans for further development
Following the initial launch of Lumigator as an MVP (Minimum Viable Product) for selecting the appropriate model, numerous other functions are planned for the project. A more detailed article from Mozilla.ai lists nine areas in which the project is intended to help with the integration and management of LLMs.
These include installation and seamless integration into projects, fine-tuning, real-time monitoring for experiments and an extended comparison of models with the same data set. The application will receive both a UI and a software development kit (SDK).
The launch of the closed alpha of Lumigator is scheduled for November 2024, and general availability is already planned for January 2025. Mozilla.ai will first conduct a survey on the desired features.
Anyone interested in Lumigator can register for the newsletter. Even though the project is currently still listed as a closed alpha, there is already a public GitHub repository that offers a Jupyter notebook with a project overview, among other things.
(rme)