Mizan-Rerank-v2 — deprecated
This model is deprecated. Use the current Mizan reranker family instead.
Mizan-Rerank-v3-Turbo is the best replacement for v2: same family, same interface, but half the size (150M vs 305M), about twice the throughput, and higher on every held-out benchmark we measure. Mizan-Rerank-v3 is the larger option (306M) if you prefer it.
The current family
| Model | Parameters | Throughput | Held-out mean nDCG@10 | Fiqh nDCG@10 | Status |
|---|---|---|---|---|---|
| Mizan-Rerank-v3-Turbo | 150M | 2,487 pairs/s | 0.875 | 0.761 | recommended replacement for v2 |
| Mizan-Rerank-v3 | 306M | 1,238 pairs/s | 0.826 | 0.669 | recommended |
| Mizan-Rerank-v2 (this model) | 305M | 1,191 pairs/s | 0.685 | 0.618 | deprecated |
Held-out mean nDCG@10 is the mean over four held-out Arabic sets (MTEB NamaaMrTydi unseen subset, short adversarial, long-context adversarial, Multi-LLM); Fiqh nDCG@10 is measured on real fatwa-search queries with production candidate lists.
Migrating
The interface is identical, so migration is one line. 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 (recommended: half the size, ~2× the throughput, more accurate)
model = CrossEncoder("ALJIACHI/Mizan-Rerank-v3-Turbo", max_length=3072, trust_remote_code=True)
# Or the larger model
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.
Links
- Recommended replacement: Mizan-Rerank-v3-Turbo
- Larger model: Mizan-Rerank-v3
- Live demo (both models, with a switcher): ALJIACHI/Mizan-Rerank-V2-Demo
- License: Apache 2.0
Note for existing work
The weights are unchanged and stay online for reproducibility: existing pipelines pinned to
ALJIACHI/Mizan-Rerank-v2 keep working. This card was reduced to a redirect on 2026-10-11; the full previous
documentation (benchmarks, training configuration, usage examples) remains in the repository history, and the
model's own benchmark chart is still at chart-1.png.
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 new work, cite the replacement:
@software{Mizan_Rerank_v3_Turbo_2026,
author = {Ali Aljiachi},
title = {Mizan-Rerank-v3-Turbo: Compact Arabic Long-Context Reranker},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/ALJIACHI/Mizan-Rerank-v3-Turbo}
}