laya-ara

Arabic NLU for typed decisions — intent, NLI, and short-list ranking on laya-multilingual.

Mohammad Alkhenizan · 21 September 2026 · LinkedIn

Hugging Face · GitHub · RAG sibling: laya-ara-rag

Highlights vs laya-multilingual

Relative lift is (fine-tune − stock) / stock on the same frozen Laya templates.

Full tables: RESULTS.md. This card’s JSON: nlu_benches.json. All cards: all_cards.json.

Abstract

laya-ara is a fine-tune of convaiinnovations/laya-multilingual (mmBERT-base with a Laya typed-decision head, ~322M) for Arabic System One inference: discrete choice, binary noul, and optional ordinal score. It is not a generative language model. Training uses official Laya RLCD on two RTX 3090 GPUs. The released mix is MASSIVE-ar, OSACT4-A, and a capped XNLI-ar sample. XNLI is CC BY-NC 4.0, so these weights are research / non-commercial (NOTICE.md).

Inference

laya is the ConvAI Laya runtime (pip install laya==0.3.4). The Hub Use this model snippet passes trust_remote_code=True and then calls laya.load. After loading, call predict (or model.predict if you used the Hub snippet). The snippet also prints this example.

pip install "laya==0.3.4"
export USE_TF=0

import os
import laya

agent = laya.load("Wouze/laya-ara", token=os.environ.get("HF_TOKEN"))
out = agent.predict(
    {"message": "الحوالة ما وصلت، أبي استرجاع وإلا بنقلع"},
    {
        "queue": {
            "type": "choice",
            "instructions": "Support queue",
            "criteria": {
                "billing": "payments, refunds",
                "technical": "bugs, outages",
                "other": "none of the above",
            },
        },
        "refund": {"type": "noul", "instructions": "Asks for a refund?"},
    },
)
print(out["answers"])

Local weights: laya.load("/path/to/artifacts/laya-ar-v48"). Demo: examples/predict_triage.py. This card’s scores: nlu_benches.json.

Method

• Base. laya-multilingual (Apache-2.0). English convaiinnovations/laya is a different checkpoint.
• Objective. Official Laya RLCD (policy gradient + soft cross-entropy, group size 4), DDP, fp16.
• Mix. MASSIVE-ar train, OSACT4 Subtask A, XNLI-ar cap. Hybrid temperature calibration (choice:11+ floor 3.75).
• Protocol. Identical frozen JSONL for stock and fine-tune. Question answering is sentence selection, not span EM/F1. Retrieval transfer is top-1 among k≤12, not corpus nDCG@10.

Classification

Published stock MASSIVE-ar intent on the Laya harness is about 0.38–0.40. XNLI remains below typical AraBERT fine-tunes (~0.80). Zero-shot AJGT and OSACT-HS decrease relative to stock. TyDiQA-ar sentence selection is 0.359 → 0.413.

Short-list reranking (this card)

No MIRACL train in this mix. Metric: top-1 among ≤12. The dedicated reranker is laya-ara-rag.

All seven listwise files improve. Pairwise Wikipedia/BM25 relevance can still favor stock.

Related models

Contact

Licensing, evaluation access, or collaboration: Mohammad Alkhenizan on LinkedIn.

Limitations

No token-level NER or span-extraction head. Offensive-language F1 is a research score, not a moderation guarantee. Diglossia is unmeasured beyond MASSIVE (ar-SA MSA) and OSACT tweets. Citations: citations.bib.

@misc{alkhenizan2026layaara,
  title        = {laya-ara: Arabic typed-decision fine-tuning of laya-multilingual},
  author       = {Alkhenizan, Mohammad},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/Wouze/laya-ara}}
}