Mizan-Rerank-v1

A compact open-source cross-encoder for reranking Arabic long texts, built for maximum accuracy per parameter.

A newer version is available: Mizan-Rerank-v2

Mizan-Rerank-v2 is the direct successor to this model and is the recommended choice for new projects.

v2 is a cross-encoder fine-tuned from Alibaba-NLP/gte-multilingual-reranker-base on 1.2M+ Arabic query–document pairs. It delivers substantially better ranking quality across every Arabic benchmark we evaluate, particularly on long documents, and it outperforms models nearly twice its size.

Use v1 when you are tightly constrained on VRAM or latency (149M parameters, ~1 GB, ~0.1 s per batch on an RTX 4090) and need the smallest viable Arabic reranker. Use v2 in every other case — especially for RAG pipelines, long-document retrieval, and domain-specific search where ranking quality is the priority.

v1 vs. v2 at a glance

	Mizan-Rerank-v1	Mizan-Rerank-v2
Parameters	149M	305M
Base model	answerdotai/ModernBERT-base	Alibaba-NLP/gte-multilingual-reranker-base
Max sequence length	8192	8192
Position embeddings	ModernBERT RoPE	RoPE with NTK scaling (factor 8.0)
Training pairs	— (741M tokens)	1,199,634 query–document pairs
Reranking (ndcg@10)	0.9986	1.0000
Triplet (ndcg@10)	0.9955	0.9993
MIRACL long docs (ndcg@10)	0.7370	0.8091
WikiQA (ndcg@10)	0.7739	0.8258
MedQA (ndcg@10)	0.5502	0.6775
Best for	minimal footprint	maximum accuracy

Migrating is a one-line change — see Migrating to v2 below.

Overview

Mizan-Rerank-v1 is an open-source reranking model based on the ModernBERT architecture, designed for reordering search results over Arabic text. At only 149 million parameters it targets the efficiency end of the spectrum: it runs comfortably in about 1 GB of memory while still handling documents up to 8192 tokens.

Key Features

Lightweight: 149M parameters, roughly half to a quarter the size of common multilingual rerankers (278M–568M)
Long text processing: Accepts up to 8192 tokens
Fast inference: Around 0.1 s response time on a single RTX 4090
Arabic language optimization: Trained specifically for Arabic morphology and context
Low memory footprint: ~1 GB VRAM, making it viable on modest GPUs and CPU-bound deployments

Performance Benchmarks

Hardware Performance (RTX 4090 24GB)

Model	Parameters	RAM Usage	Response Time
Mizan-Rerank-v1	149M	1 GB	0.1 s
jina-reranker-v2-base-multilingual	278M	2.5 GB	0.2 s
bge-reranker-v2-m3	568M	4 GB	0.3 s

Efficiency remains v1's primary advantage. If your bottleneck is memory or latency rather than ranking quality, v1 is still the smallest model in the Mizan family.

Ranking Quality (ndcg@10)

The table below comes from the unified evaluation run used for the v2 release, so all models are measured under identical conditions and the numbers are directly comparable.

Model	Parameters	Reranking	Triplet	MIRACL (Long Docs)	WikiQA	MedQA
ALJIACHI/Mizan-Rerank-v2	305M	1.0000	0.9993	0.8091	0.8258	0.6775
BAAI/bge-reranker-v2-m3	568M	1.0000	0.9998	0.7231	0.8669	0.6584
Alibaba-NLP/gte-multilingual-reranker-base	305M	1.0000	0.9991	0.7539	0.8275	0.6648
ALJIACHI/Mizan-Rerank-v1	149M	0.9986	0.9955	0.7370	0.7739	0.5502

v1 stays competitive with models 2–4× its size on the saturated reranking and triplet sets. The gap widens on long-document MIRACL and on domain-specific sets such as MedQA, which is exactly what v2 was built to close.

Note on earlier figures: the original v1 model card reported a MIRACL ndcg@10 of 0.8865, measured on a different evaluation subset and protocol. The table above supersedes it and should be used for any comparison against v2 or other rerankers.

Training Methodology

Mizan-Rerank-v1 was trained on a corpus of 741,159,981 tokens drawn from:

Authentic Arabic open-source datasets
Manually curated and processed text
Purpose-generated synthetic data

How It Works

Query reception: The model receives a user query alongside candidate passages
Content analysis: It jointly encodes each query–passage pair and analyzes their semantic relationship
Relevance scoring: It assigns a relevance score to each passage
Reranking: Passages are sorted by descending score

Usage

Using Transformers

from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch

model = AutoModelForSequenceClassification.from_pretrained("ALJIACHI/Mizan-Rerank-v1")
tokenizer = AutoTokenizer.from_pretrained("ALJIACHI/Mizan-Rerank-v1")
model.eval()

def get_relevance_score(query, passage):
    inputs = tokenizer(
        query, passage,
        return_tensors="pt", padding=True, truncation=True, max_length=8192,
    )
    with torch.no_grad():
        logits = model(**inputs).logits
    # Apply sigmoid to map the raw logit into a 0–1 relevance probability
    return torch.sigmoid(logits).item()

query = "ما هو تفسير الآية وجعلنا من الماء كل شيء حي"
passages = [
    "تعني الآية أن الماء هو عنصر أساسي في حياة جميع الكائنات الحية، وهو ضروري لاستمرار الحياة.",
    "تم اكتشاف كواكب خارج المجموعة الشمسية تحتوي على مياه متجمدة.",
    "تحدث القرآن الكريم عن البرق والرعد في عدة مواضع مختلفة.",
]

scores = [(p, get_relevance_score(query, p)) for p in passages]
reranked = sorted(scores, key=lambda x: x[1], reverse=True)

for passage, score in reranked:
    print(f"Score: {score:.4f} | {passage}")

Using Sentence Transformers

pip install -U sentence-transformers

from sentence_transformers import CrossEncoder

model = CrossEncoder("ALJIACHI/Mizan-Rerank-v1", max_length=8192)

ranks = model.rank(
    "ما هي فوائد فيتامين د؟",
    [
        "يساعد فيتامين د في تعزيز صحة العظام وتقوية الجهاز المناعي، كما يلعب دوراً مهماً في امتصاص الكالسيوم.",
        "يستخدم فيتامين د في بعض الصناعات الغذائية كمادة حافظة.",
        "يمكن الحصول على فيتامين د من خلال التعرض لأشعة الشمس أو تناول مكملات غذائية.",
    ],
)
print(ranks)

Migrating to v2

The interface is identical; only the model ID changes, plus trust_remote_code=True for v2's custom architecture:

from sentence_transformers import CrossEncoder

# Before
model = CrossEncoder("ALJIACHI/Mizan-Rerank-v1", max_length=8192)

# After
model = CrossEncoder("ALJIACHI/Mizan-Rerank-v2", max_length=8192, trust_remote_code=True)

Scores are not calibrated across versions, so if your pipeline uses an absolute score threshold to filter candidates, re-tune that threshold after switching. Full details and a live demo are on the v2 model card.

Practical Examples

Example 1

السؤال: ما هو القانون الجديد بشأن الضرائب في 2024؟

النص	الدرجة
نشرت الجريدة الرسمية قانوناً جديداً في 2024 ينص على زيادة الضرائب على الشركات الكبرى بنسبة 5%	0.9989
الضرائب تعد مصدراً مهماً للدخل القومي وتختلف نسبتها من دولة إلى أخرى.	0.0001
افتتحت الحكومة مشروعاً جديداً للطاقة المتجددة في 2024.	0.0001

Example 2

السؤال: ما هو تفسير الآية وجعلنا من الماء كل شيء حي

النص	الدرجة
تعني الآية أن الماء هو عنصر أساسي في حياة جميع الكائنات الحية، وهو ضروري لاستمرار الحياة.	0.9996
تم اكتشاف كواكب خارج المجموعة الشمسية تحتوي على مياه متجمدة.	0.0000
تحدث القرآن الكريم عن البرق والرعد في عدة مواضع مختلفة.	0.0000

Example 3

السؤال: ما هي فوائد فيتامين د؟

النص	الدرجة
يساعد فيتامين د في تعزيز صحة العظام وتقوية الجهاز المناعي، كما يلعب دوراً مهماً في امتصاص الكالسيوم.	0.9991
يستخدم فيتامين د في بعض الصناعات الغذائية كمادة حافظة.	0.9941
يمكن الحصول على فيتامين د من خلال التعرض لأشعة الشمس أو تناول مكملات غذائية.	0.9938

Example 3 illustrates a known v1 weakness: it separates relevant from irrelevant text reliably, but its scores are less discriminative among passages that merely mention the query topic. v2 handles this case better.

Model Details

Model Type: Cross Encoder
Base Model: answerdotai/ModernBERT-base
Parameters: 149M
Maximum Sequence Length: 8192 tokens
Number of Output Labels: 1
Language: Arabic (ar), English (en)
License: Apache 2.0
Successor: ALJIACHI/Mizan-Rerank-v2

Applications

Specialized Arabic search engines
Archiving systems and digital libraries
RAG (Retrieval-Augmented Generation) pipelines
Conversational AI applications
E-learning platforms
Information retrieval systems

For any of the above where ranking quality matters more than footprint, start with v2.

Citation

If you use Mizan-Rerank-v1 in your research, please cite:

@software{Mizan_Rerank_v1_2025,
  author = {Ali Aljiachi},
  title = {Mizan-Rerank-v1: A Revolutionary Arabic Text Reranking Model},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/ALJIACHI/Mizan-Rerank-v1}
}

For the successor model:

@software{Mizan_Rerank_v2_2026,
  author = {Ali Aljiachi},
  title = {Mizan-Rerank-v2: Arabic Long-Context Text Reranking Model},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/ALJIACHI/Mizan-Rerank-v2}
}

Base model:

@misc{modernbert,
      title={Smarter, Better, Faster, Longer: A Modern Bidirectional Encoder for Fast, Memory Efficient, and Long Context Finetuning and Inference},
      author={Benjamin Warner and Antoine Chaffin and Benjamin Clavié and Orion Weller and Oskar Hallström and Said Taghadouini and Alexis Gallagher and Raja Biswas and Faisal Ladhak and Tom Aarsen and Nathan Cooper and Griffin Adams and Jeremy Howard and Iacopo Poli},
      year={2024},
      eprint={2412.13663},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2412.13663},
}