aaronalexS/daylens-laya-onnx

Laya (typed-decisions) — ONNX export for Daylens

This is an ONNX (fp32) conversion of the typed-decisions checkpoint of convaiinnovations/laya, made for the Daylens desktop app, which runs it locally with onnxruntime-node to label on-screen activity.

All credit for the model goes to Convai Innovations. The original model is licensed under Apache-2.0; this derivative is distributed under the same license (see LICENSE).

Changes from the original

• Exported to ONNX (fp32) with dynamic sequence length and a traceable decision head (question/option marker positions are passed as inputs). Weights are unchanged.
• laya-meta.json holds the special-token ids and option-marker metadata the Daylens TypeScript runtime needs to format inputs the same way as the official Python code.
• Parity with the original PyTorch model: max |logit difference| 0.00005; 96/96 matching decisions on the Daylens golden set.

Files

File	Size	SHA-256
laya.onnx	1 686 012 251 B	bbd684549c90cab727e43fd1d6c9458cf6e791a6af5913110746c98ba4253ad8
tokenizer.json	3 583 228 B	6c8aaa9a542084f2457eab775d4eeb51f92a70c0fd9de28d5edb0ddec3c08d30
tokenizer_config.json	337 B	08d4cf3ac4dca381759441b85b91a6d40e688471dcd33d15d6649eb0a9a854d1
laya-meta.json	519 B	429b6917f0f855257f18500f163b6a01860fa6d53a0747f4e0f4ebad103a34e9

Intended use

Local, on-device classification of short screen-context strings (app name, window title, redacted OCR text) inside Daylens. For general use, prefer the original model.