Wouze/laya-ara-rag

laya-ara-rag

Arabic short-list reranker — passage relevance and k≤12 ranking on `laya-multilingual`.

<p align="center"> <img src="assets/logo.jpg" alt="laya-ara-rag" width="280"> </p>

Mohammad Alkhenizan · 21 September 2026 · LinkedIn


`Hugging Face` · `GitHub` · NLU: `laya-ara`

Highlights vs laya-multilingual

Relative lift is (fine-tune − stock) / stock. Metric is top-1 among ≤12 candidates, not corpus nDCG@10.

Mintaka entity ranking does not improve (0.403 predecessor vs 0.388). Full tables: `RESULTS.md`. This card’s JSON: `rag_all_benches.json`. All cards: `all_cards.json`.

Abstract

laya-ara-rag is a fine-tune of `convaiinnovations/laya-multilingual` (mmBERT-base with a Laya typed-decision head, ~322M) for Arabic pairwise relevance (noul) and listwise reranking (choice). It is not generative and not a first-stage retriever. Training uses official Laya RLCD on two RTX 3090 GPUs.

The mix continues from a MIRACL-ar Wikipedia specialist, then adds in-house Arabic retrieval logs plus a MIRACL-ar train replay and a small MASSIVE-ar cap. XNLI is withheld. Evaluation is top-1 among a shortlist of k≤12.

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-rag", token=os.environ.get("HF_TOKEN"))
out = agent.predict(
    {"query": "ما حكم الوضوء قبل قراءة القرآن؟"},
    {
        "passage": {
            "type": "choice",
            "instructions": "Which passage is most relevant to the query?",
            "criteria": {
                "a": "الوضوء شرط للصلاة لا للقراءة عند جمهور الفقهاء.",
                "b": "زكاة الفطر تجب على كل مسلم قبل صلاة العيد.",
                "c": "صيام عاشوراء سنة مؤكدة عند الحنابلة.",
            },
        },
        "relevant": {
            "type": "noul",
            "instructions": "Is passage A relevant to the query?",
        },
    },
)
print(out["answers"])

The example passages are illustrative. Demo: `examples/predict_rerank.py`. Scores: `rag_all_benches.json`.

Method

• —Base. laya-multilingual (Apache-2.0). Init from the MIRACL-ar specialist, not from laya-ara.
• —Objective. Official Laya RLCD, DDP, fp16. One epoch, 35,591 items.
• —Public mix. MIRACL-ar train pairwise + listwise; MASSIVE-ar cap 2,000.
• —In-house mix. Arabic retrieval logs (not redistributed): teacher-distilled pairs with a minimum score gap, listwise gold = teacher rank-1, sealed hold-out. See `DATA.md`.
• —Protocol. Frozen JSONL. Pairwise snippets ~800 characters; listwise ~400 so k = 12 fits max_len 1024.

Public listwise rerank (k ≤ 12)

Chance ≈ 0.08–0.11. MIRACL-ar train is in the mix (dev is in-family). Other rows are transfer.

Mintaka listwise favors the Wikipedia-only predecessor (0.403). PublicHealthQA n is small.

Public pairwise relevance

One gold + one negative. Chance = 0.50.

Listwise transfer does not imply pairwise transfer. Entity and product pairwise still favor the NLU card.

In-domain exam

Sealed hold-out on the in-house Arabic retrieval stack (queries and passages not in this repository).

Teacher-hold is distillation agreement, not clicks. The 21-query human listwise set is too small to headline.

Related models

Contact

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

Limitations

Not a first-stage retriever and not official MIRACL nDCG@10. No span extraction or token NER. Mintaka / MLQA / XPQA pairwise can drop versus laya-ara.

@misc{alkhenizan2026layaararag,
  title        = {laya-ara-rag: Arabic short-list reranking with laya-multilingual},
  author       = {Alkhenizan, Mohammad},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/Wouze/laya-ara-rag}}
}