convaiinnovations/laya-multilingual — new model trending #30 on Hugging Face
ConvAI published Laya Multilingual, a 322M non-autoregressive model answering typed questions across 100-plus languages.
Laya Multilingual is a 322-million-parameter non-autoregressive decision model from convaiinnovations, built on mmBERT-base with a 1024-token context. It takes a state plus typed questions and returns scored answers in one forward pass, intended for 100-plus languages rather than English. On MASSIVE intent classification it reaches 0.366 macro accuracy versus 0.227 for the English Laya checkpoint, and batched inference is faster. The authors say probabilities ship uncalibrated and English results are weaker than the English-only checkpoint.
- 322M params on mmBERT-base with 1024-token context
- MASSIVE macro accuracy 0.366 versus 0.227 for English Laya
- XNLI non-English average 0.731 versus 0.521
- Ships over-confident; authors recommend temperature calibration
- Router in laya 0.3.7 keeps both checkpoints resident
- Products
- laya
- Organizations
- convaiinnovations
From the original article
5 min read at huggingface.cotext-classification · 0 downloads · 194 likes · license apache-2. 0
By convaiinnovations · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.