Add Laya: 421M RLCD decision model + demo Space (trained)
Adds convaiinnovations/laya + laya-demo to the trained category.
Why it belongs: 421M params (ModernBERT-large fully fine-tuned + decision head trained from scratch), trained with RLCD against proper scoring rules, choice/score/noul with calibrated probabilities, all questions answered in one pass. Apache-2.0 weights at https://huggingface.co/convaiinnovations/laya and a 7-tab Gradio demo at https://huggingface.co/spaces/convaiinnovations/laya-demo (triage, email, guardrails, RAG, moderation, routing, playground).
Metrics snapshot: model 2 likes (2026-09-18), demo Space 1 like. No launch post on X found, so no X metrics; sorts on Hub likes only. Dates verified from Hub metadata (space created 2026-09-18, model published 2026-09-18).
Change is a single 4-line ITEMS entry; data-only, no renderer changes.
Merged — thanks! Laya is in the tracker under trained.
I verified the repo (NandhaKishorM/laya, Apache-2.0), the Hub model (convaiinnovations/laya) and the demo Space all exist, and corrected two things while merging:
- Metrics. The PR had 8 stars / 2 model likes; the live figures are 163 stars / 108 model likes. The entry now carries the current snapshot.
- Description length. Trimmed from 326 to 259 chars to match the card ceiling used across the tracker.
The demo-Space like count was dropped in favour of the model likes, since that's the field the other entries use.
One note: there's no launch post on X, so the card sorts on GitHub stars and Hub likes only — with 163 stars it still lands high in the trained category.