Laya Sentiment Multilingual

Binary sentiment classification for English, Spanish, and German

🧠 Laya decision model 📦 322M params · mmBERT-base 🍎 CoreML FP16 · ANE 📜 Apache 2.0

88.0%

Holdout accuracy

0.025

ECE (calibrated)

16 ms

Steady-state latency

Marginal cost

Interactive demo — curated examples

This static Space shows real predictions from the 300-example holdout. Select an example to see the model's output, calibrated confidence, and delegation decision.

Language

—

Prediction

—

Calibrated confidence

—

Decision

—

Delegation policy: if calibrated confidence < 0.75, the harness delegates to cloud (DeepSeek V4-Flash). This is the same mechanism that runs on-device at 16 ms in production.

Accuracy by language (fine-tuned)

92.0%English

85.0%Spanish

87.0%German

100 examples per language. Differences are not statistically distinguishable (pairwise z-tests, all p > 0.1). The model holds up across EN/ES/DE within the tested distribution.

Production pipeline (Hybrid Harness V5)

1. InputText (EN/ES/DE)

2. CoreMLANE · FP16 · 16 ms

3. CalibrationTemperature T=0.9

4. Policyconf < 0.75 ?

5. OutputLocal or cloud

The local model handles ~90.3% of traffic. The remaining ~9.7% is delegated to DeepSeek V4-Flash when the calibrated confidence indicates uncertainty.

Model evolution (V3 → V5)

V3 (routing, 4 classes): 93.5% internal, but 46.7% external (Tobi-Bueck). Root cause: organizational labels, not semantic ground truth.

V4 (pivot to sentiment): Switched to a publicly verifiable task. +32.3 points of real gain (55.7% zero-shot → 88.0% fine-tuned).

V5 (production-ready): CoreML FP16, temperature scaling, delegation policy, end-to-end validated.