Convai Innovations releases laya-typed-decisions, a 0.4B ModernBERT decision model
An Apache-2.0 encoder fine-tuned on four synthetic workflows. The card reports 0.766 accuracy and warns its probabilities are over-confident.
Convai Innovations published laya-typed-decisions on Hugging Face on 18 September 2026, by repository creation date. The card calls it a non-autoregressive System 1 decision model for text classification, part of the Laya family. It is fine-tuned on four synthetic workflows: agent-trace observability, customer service, invoice processing and security incidents.
It has about 0.4B parameters, is built on ModernBERT-large via convaiinnovations/laya, and is licensed Apache-2.0. Context is 1,024 tokens and the card says English only. Docs, GitHub, PyPI and a benchmark dataset are linked.
On 400 test cases, 2,000 decisions, the card reports accuracy 0.766, Brier 0.062, ECE 0.213 and score MAE 0.242. It lists Jev 1.13.0 at accuracy 0.727, with Jev figures taken from a third party rather than measured by the authors, and calls the comparison indicative only. By workflow, invoice processing is highest at 0.804 and agent-trace observability lowest at 0.730.
The card is blunt about limits. It is a specialist and may perform poorly outside its four workflows. Its probabilities are over-confident and should be recalibrated on your own held-out data. Choice questions should stay under about 20 options.
Small, specific, and honest about calibration.
Sources
Editor's first pass (AI-assisted)
- Real 4: Weights, PyPI package, docs and GitHub are listed on the card.
- Useful 3: Narrow: four synthetic workflows, English, 1,024-token context.
- Notable 2.5: Several community repos on the Hub reuse its name.
- Novel 3: A non-autoregressive encoder rather than an LLM fine-tune.
Comments