The multilingual Aurora-m-instruct model with 15 billion parameters has been released

Developers presented Aurora-m-instruct — a specialized language model with 15 billion parameters, focused on executing instructions, the release is aimed at developing tools for safe interaction with AI in a multilingual environment. The basis of the architecture is the Starcoderplus base, refined to support a wide range of language tasks, the model training process included processing 1.5 trillion tokens of data, special attention was paid to security during dataset preparation: the training sample integrates the results of red-teaming research, which allows the model to more effectively recognize and block malicious requests. The model is available for download on the Hugging Face platform, its technical characteristics allow it to be used in tasks that require high accuracy in following instructions when working with multilingual texts, the developers emphasize that the main goal of the project is to create an open tool for academic and applied research in the field of AI system security. The release of this model expands the available resources for developers working with compact but productive solutions, the use of data that has undergone red-teaming makes Aurora-m-instruct suitable for integration into corporate systems where control over model behavior and minimization of risks of generating undesirable content are critically important.