DigiNews

Tech Watch by Johan Denoyer

← Back to articles

Laya, un modèle de décision multilingue non-autorégressif pour la classification typée

Quality: 8/10 Relevance: 9/10

Summary

Laya is a multilingual, non-autoregressive decision model for typed classification. It outputs typed answers with calibrated probabilities in a single forward pass across 100+ languages, trained with RLCD to encourage honest probabilities and avoid text generation. The model family includes English, multilingual, and typed-decisions checkpoints; it uses a ModernBERT-large backbone (English) and mmBERT-base for multilingual, all under Apache 2.0, with a Router that routes by language. The project emphasizes fast inference and a route-based deployment path.

🚀 Service construit par Johan Denoyer