---
title: Laya-Hebrew Decision Model
emoji: 🎯
colorFrom: blue
colorTo: pink
sdk: gradio
sdk_version: 6.29.0
app_file: app.py
pinned: false
license: cc-by-nc-sa-4.0
short_description: Calibrated Hebrew & English decisions
python_version: "3.12"
startup_duration_timeout: 1h
---

# Laya-Hebrew — calibrated Hebrew & English decisions

[`RoeiG/laya-hebrew`](https://huggingface.co/RoeiG/laya-hebrew) is a **decision** model, not a
chatbot. Give it a **state** (the text or fields to judge) and **typed questions**, and one
forward pass returns a **calibrated probability for every answer**. It never generates text,
so there is nothing to parse and nothing to hallucinate.

| type | question | answer |
|---|---|---|
| `choice` | which of these options? | the option, a probability per option, confidence |
| `score` | where on this rubric? | the expected level, a probability per level, confidence |
| `noul` | is this claim true? | P(true) |

378M parameters (363M NeoBERT encoder + 15M head), fp16, 1,024 tokens per question. Hebrew and
English only.

## How to get good answers

1. **Compute numbers, dates, units and relations in code** and pass the result as a field (the
   *Advanced options* box). The model does not do arithmetic reliably — and it is often
   *confidently* wrong: 2.5 hours against a 2-hour limit, or 19 days against a 14-day window,
   got P(true) of 0.93–0.98.
2. **Prefer the claim form for yes/no.** "הלקוח כועס." discriminates better than
   "האם הלקוח כועס?" (gap 0.67 against 0.38 on he_bench).
3. **Describe every option in a line.** Bare labels or codes route much worse than labels with
   a one-line description.
4. **Read the probabilities, not only the top answer.** A top answer below ~0.6 means the model
   is unsure — and a confidence threshold will *not* catch the failures listed below, which are
   typically high-confidence.
5. **Ignore `act_probability`.** It comes from a head that no Hebrew checkpoint trained; it is
   left in the raw JSON and hidden from the rendered answers.

## Limitations

- Reasoning is the weak spot: hellaswag 0.447, winograd 0.607, copa 0.687. Multi-step
  inferences ("A is taller than B, B is taller than C: who is shortest?") often fail.
- Numeric, date and unit rules are unreliable and often confidently wrong. Compute them in code.
- Irony and sarcasm are read literally ("וואו, שירות מדהים…" is scored as positive, at 0.97).
- Routing leans on keywords: a strong keyword can outweigh the option descriptions.
- Scales are weak: sentiment and tone are ~0.45 accuracy, and tone probabilities are
  overconfident (valence Brier 0.53).

Held-out evaluation: MASSIVE-he 20 intents 0.806, he_bench accuracy 0.595 (Brier 0.499,
ECE 0.125), Hebrew BoolQ 0.838, Belebele-he 0.767 (chance 0.25).

## Implementation notes

This Space vendors **Laya 0.3.7** at commit `010bace` in `laya/`, with
[`neobert.patch`](https://huggingface.co/RoeiG/laya-hebrew/blob/main/neobert.patch) applied:

- NeoBERT's modeling code lives on the Hub, so `trust_remote_code=True` is passed through.
- NeoBERT computes its rotary tables as non-persistent buffers in `__init__`; under
  transformers 5 the model is built on the meta device, so they come back as uninitialised
  memory and **every forward pass returns NaN**. The patch recomputes them.
- The encoder config pins `bfloat16`; the patch keeps the encoder in **fp32**.

Attention is SDPA, the model is loaded once at module scope and runs on ZeroGPU
(`@spaces.GPU`).

## License

**CC-BY-NC-SA-4.0 — non-commercial use only.** Several training sources are non-commercial or
research-only (ANLI, Yelp, AG News, Yahoo Answers, RACE, and the NLLB translation model).

- Encoder: [`dicta-il/neodictabert-bilingual`](https://huggingface.co/dicta-il/neodictabert-bilingual)
  (NeoBERT, by Dicta) — CC-BY-4.0.
- Laya architecture, training method and runtime: [Laya](https://github.com/NandhaKishorM/laya)
  (Apache-2.0). No weights come from Laya's published checkpoints.