convaiinnovations/laya — new model trending #30 on Hugging Face
Convaiinnovations releases Laya, an open-source 421M-parameter non-autoregressive decision model for calibrated email triage, phishing detection, and routing.
Convaiinnovations published a fine-tuned checkpoint of Laya, an open-source (Apache 2.0) non-autoregressive System 1 decision model trending #30 on Hugging Face. It combines a fully fine-tuned ModernBERT-large backbone (395M) with a from-scratch decision head for 421M total parameters, returning typed answers with calibrated probabilities instead of generated text. Training used RLCD with proper scoring rules and TD(lambda=1.0) trajectory modeling on fully human-annotated data. Reported results include 38.4 ms single-question latency, 83.8% in-task macro accuracy, and 73.2% accuracy on email triage and phishing detection.
- Fine-tuned Laya checkpoint trending #30 on Hugging Face under an Apache 2.0 license.
- ModernBERT-large (395M) backbone plus a 2-layer decision head, 421M parameters total.
- RLCD and TD(lambda=1.0) training yield calibrated probabilities without text generation or hallucination.
- Reports 38.4 ms single-question latency and 83.8% in-task macro accuracy across task families.
- Email triage hits 73.2% accuracy for phishing/scam detection and 99.1% for intent routing.
- Tags
- calibrationemail-triagehugging-facemodernbertnon-autoregressiveopen-weightsphishing-detectionrlcd
- Vendors
- Convaiinnovations
- Products
- LayaModernBERT-large
- Organizations
- Convaiinnovations
- AI models
- Laya
From the original article
4 min read at huggingface.coreinforcement-learning · 0 downloads · 144 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.