Mizan-Rerank-v2

A cross-encoder model for reranking Arabic long texts, fine-tuned from Alibaba-NLP/gte-multilingual-reranker-base.

---

Deprecated: use Mizan-Rerank-v3

This model is deprecated and no longer maintained. Use Mizan-Rerank-v3 instead.

v3 is a 306M-parameter cross-encoder, fine-tuned from the same Alibaba-NLP/gte-multilingual-reranker-base base as v2 but on 71k Arabic listwise groups whose hard negatives are adversarial traps: negated rulings, swapped entities, shifted numbers and dates, exceptions applied to the wrong case. It scores higher than v2 on every held-out benchmark we measure, and it runs about twice as fast as bge-reranker-v2-m3.

v3 is trained from the base model rather than from v2, so it is a replacement rather than a continuation. v3 covers the use cases v2 was built for, so there is no reason to start a new project on v2. This card is kept online for reproducibility of existing work.

v2 vs. v3 at a glance

	Mizan-Rerank-v2	Mizan-Rerank-v3
Status	deprecated	recommended
Parameters	305M	306M
Base model	Alibaba-NLP/gte-multilingual-reranker-base	Alibaba-NLP/gte-multilingual-reranker-base
Trained from	the gte base	the gte base (not from v2)
Training data	1,199,634 Arabic query–document pairs	71,044 Arabic listwise groups with adversarial hard negatives
Held-out Arabic benchmarks	baseline	higher on every set
Best for	—	RAG, long-document retrieval, domain-specific Arabic search

Migrating is a one-line change — see Migrating to v3 below.

---

Overview

Mizan-Rerank-v2 is a cross-encoder reranking model based on Alibaba-NLP/gte-multilingual-reranker-base, fine-tuned for Arabic text reranking. It handles long documents (up to 8192 tokens) and was the state of the art in the Mizan family when it was released. It is now superseded by v3.

Legacy demo

The v2 demo space is still online at ALJIACHI/Mizan-Rerank-V2-Demo. It serves v2; use v3 for new work.

Key Features

• Long Document Support: Handles up to 8192 tokens using RoPE position embeddings with NTK scaling
• Compact for its quality: 305M parameters, roughly half the size of BAAI/bge-reranker-v2-m3 (568M)
• Arabic Language Optimization: Fine-tuned on 1.2M+ Arabic query-document pairs from diverse sources

Performance Benchmarks

Reranking Evaluation (ndcg@10)

The table below comes from the evaluation run used for the v2 release, so all models are measured under identical conditions and the numbers are directly comparable. It predates v3.

Model	Parameters	Reranking	Triplet	MIRACL (Long Docs)	WikiQA	MedQA
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

Key Improvements over Base Model

Benchmark	Base Model	Mizan-Rerank-v2	Improvement
Reranking	1.0000	1.0000	--
Triplet	0.9991	0.9993	+0.0002
MIRACL (Long Docs)	0.7539	0.8091	+0.0552
WikiQA	0.8275	0.8258	-0.0017
MedQA	0.6648	0.6775	+0.0127

Key Improvements over BAAI/bge-reranker-v2-m3

Benchmark	bge-reranker-v2-m3	Mizan-Rerank-v2	Improvement
Reranking	1.0000	1.0000	--
Triplet	0.9998	0.9993	-0.0005
MIRACL (Long Docs)	0.7231	0.8091	+0.0860
WikiQA	0.8669	0.8258	-0.0411
MedQA	0.6584	0.6775	+0.0191

Model Details

• Model Type: Cross Encoder
• Base Model: Alibaba-NLP/gte-multilingual-reranker-base
• Architecture: NewForSequenceClassification (12 layers, 768 hidden, 12 heads)
• Maximum Sequence Length: 8192 tokens
• Position Embeddings: RoPE with NTK scaling (factor 8.0)
• Number of Output Labels: 1
• Language: Arabic (ar), English (en)
• License: Apache 2.0
• Status: deprecated
• Replacement: ALJIACHI/Mizan-Rerank-v3

Usage

Using Sentence Transformers

pip install -U sentence-transformers

from sentence_transformers import CrossEncoder

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

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

scores = model.predict(pairs)
print(scores)
# High score for the relevant passage, low scores for irrelevant ones

# Or rank documents for a query
ranks = model.rank(
    "ما هو تفسير الآية وجعلنا من الماء كل شيء حي",
    [
        "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.",
    ]
)
print(ranks)
# [{'corpus_id': 0, 'score': ...}, {'corpus_id': 1, 'score': ...}, ...]

Using Transformers Directly

from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch

model = AutoModelForSequenceClassification.from_pretrained(
    "ALJIACHI/Mizan-Rerank-v2",
    trust_remote_code=True,
    torch_dtype=torch.float16,
)
tokenizer = AutoTokenizer.from_pretrained("ALJIACHI/Mizan-Rerank-v2")

def get_relevance_score(query, passage):
    inputs = tokenizer(query, passage, return_tensors="pt", padding=True, truncation=True, max_length=8192)
    with torch.no_grad():
        outputs = model(**inputs)
    return torch.sigmoid(outputs.logits).item()

query = "ما هي فوائد فيتامين د؟"
passages = [
    "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.",
    "أطلقت وزارة الزراعة حملة وطنية لزيادة الوعي بأهمية الزراعة العضوية.",
]

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[:80]}...")

Migrating to v3

The interface is identical; only the model ID changes. Both models need trust_remote_code=True:

from sentence_transformers import CrossEncoder

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

# After
model = CrossEncoder("ALJIACHI/Mizan-Rerank-v3", max_length=3072, 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, benchmarks and a usage guide are on the v3 model card.

Training Details

Training Data

Trained on 1,199,634 query-document pairs from diverse Arabic sources

Training Configuration

Parameter	Value
Base Model	Alibaba-NLP/gte-multilingual-reranker-base
Max Sequence Length	8192
Batch Size	2
Gradient Accumulation Steps	16
Effective Batch Size	32
Learning Rate	5e-7
LR Scheduler	Cosine
Warmup Ratio	0.1
Precision	FP16
Gradient Checkpointing	Enabled
Loss Function	BinaryCrossEntropyLoss (pos_weight=1.24)

Applications

• Arabic search engines and information retrieval systems
• RAG (Retrieval-Augmented Generation) pipelines
• Islamic text search and jurisprudence Q&A
• Digital library and archive search
• Long-document Arabic content analysis
• E-learning platforms with Arabic content

For all of the above, use v3. This model is deprecated and is documented here only for reproducibility of existing work.

Framework Versions

• Python: 3.10.14
• Sentence Transformers: 5.4.1
• Transformers: 4.55.4
• PyTorch: 2.8.0+cu126
• Accelerate: 1.10.0
• Datasets: 3.5.0
• Tokenizers: 0.21.0

Citation

@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}
}

For the recommended model:

@software{Mizan_Rerank_v3_2026,
  author = {Ali Aljiachi},
  title = {Mizan-Rerank-v3: Adversarially Trained Arabic Long-Context Reranker},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/ALJIACHI/Mizan-Rerank-v3}
}

@inproceedings{reimers-2019-sentence-bert,
    title = "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks",
    author = "Reimers, Nils and Gurevych, Iryna",
    booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing",
    month = "11",
    year = "2019",
    publisher = "Association for Computational Linguistics",
    url = "https://arxiv.org/abs/1908.10084",
}

License

Released under the Apache 2.0 License.