AI Briefinglaya-multilingual: Text classification in 100 languages, 2x faster with 322M parameters
convaiinnovations/laya-multilingual
About the project
When receiving various inputs and questions such as text, emails, and tickets, it immediately returns probability-based typed answers. Unlike generative models, there is no text generation process, so parsing errors or hallucinations do not occur. It processes with a single forward pass, resulting in short response latency and high predictability.
A 322M parameter model based on the mmBERT-base encoder, supporting over 100 languages. Accuracy is significantly improved for non-English languages compared to English-only models, showing usable performance especially in Arabic, Hindi, Korean, and others. Processing speed is approximately 2x faster than English models, at 72.3ms for 10 questions.
Through the Router feature, it can automatically branch between English and multilingual models based on the input language. Pre-trained checkpoints can be kept resident in memory to eliminate loading delays during language switching. It is available under the Apache 2.0 license for commercial use and can be efficiently deployed in low-resource environments.
convaiinnovations/laya-multilingual
The original page has no description.
text-classification
This introduction was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report errors, attribution issues, or removal requests via Contact.