Repo

laya

Evaluates choice, score, and boolean questions over documents or JSON in one pass, then routes requests to the best checkpoint by detected language/script.

GitHub ↗Sep 18, 2026

  • nlp
  • classification
  • multilingual
  • calibration
Comprehensive comparative dashboard: Laya vs Jev accuracy across benchmarks, headline claims, coverage
Comprehensive comparative dashboard: Laya vs Jev accuracy across benchmarks, headline claims, coverage

What it is

Single-pass typed decision engine for text/JSON that routes across English and multilingual checkpoints. Returns labels, scores, or calibrated booleans without generation. Answer typed questions over text, emails, tickets, or JSON in one forward pass, with automatic routing to English, multilingual, or fine-tuned decision checkpoints and confidence-bearing outputs.

Key results

  • 32.8 ms for 1 question; 72.3 ms for 10 on a T4
  • Router reaches 45/51 usable languages; English-only model gets 23/51
  • Typed-decisions checkpoint scores 0.766 vs published Jev 0.727

Shows that many production decision workflows can be handled by non-generative, single-pass classifiers with useful confidence estimates.

One useful ML email a week.

Worth your time, not all of your time.

We track new ML work across papers, code, and data, then keep only the pieces worth a closer look. Selected, not collected.

Papers.Data.Code

The Weekly ML Signal

Selected, not collected.

One useful ML email a week: the papers, repos and datasets worth your attention.

Worth your time, not all of your time.