Intiqa
Puts the passages that answer a question about Islam first. It reranks 20 search results in about 0.37 s on an 8-core CPU.
It runs in Sirathic, a free Islamic answering engine, between search and the answer model.
pip install sentence-transformers
from sentence_transformers import CrossEncoder
intiqa = CrossEncoder("sirathic/intiqa", max_length=256)
passages = [
"Combining Maghrib and Isha because of rain is allowed by the Shafi'i, Maliki and Hanbali schools...",
"The times of the five daily prayers are...",
]
for hit in intiqa.rank("Can I combine Maghrib and Isha when it's raining?", passages):
print(hit["corpus_id"], hit["score"])
rank returns the passages best first. A score is a raw logit: higher means more relevant. Compare scores only within one question.
Benchmarks
293 questions about Islam (English 90, Arabic 88, Urdu 52, Indonesian 48, Turkish 9, French 9), each with about 20 passages from Quran, hadith, fatwa and fiqh search, graded by Claude Opus. nDCG@10, and how often a passage that directly answers the question comes first.
| Model | 256 tokens | Top-1 | 512 tokens | Top-1 |
|---|---|---|---|---|
| Search alone, no reranker | 0.625 | 44.5% | 0.625 | 44.5% |
| mMiniLM-L12 base (before training) | 0.691 | 53.5% | 0.698 | 56.7% |
| gte-multilingual-reranker-base | 0.702 | 55.1% | 0.719 | 55.1% |
| Qwen3-Reranker-0.6B | 0.740 | 56.3% | 0.774 | 64.1% |
| Intiqa | 0.746 | 58.0% | 0.786 | 61.2% |
| Our fine-tuned Qwen3-Reranker-0.6B (Intiqa's teacher) | 0.774 | 56.7% | 0.817 | 64.9% |
Time to rerank 20 passages, p95, bfloat16, one batch, on AWS c7i (Intel Sapphire Rapids):
| Model | 256 tokens, 4 cores | 256 tokens, 8 cores | 128 tokens, 8 cores |
|---|---|---|---|
| Intiqa | 0.69 s | 0.37 s | 0.17 s |
| gte-multilingual-reranker-base | 2.1 s | 1.3 s | 0.71 s |
| Qwen3-Reranker-0.6B | 10 s | 5.5 s | 4.1 s |
These times need a CPU with bfloat16 support, such as AMD Zen 4 or Intel Sapphire Rapids. On other CPUs, fp32 takes about twice as long. ONNX int8 was never faster and scored about 0.02 lower.
Training
Intiqa learned from its teacher, our fine-tuned Qwen3-Reranker-0.6B. The teacher scored 853,320 question and passage pairs from 50,000 questions sent to Islamic fatwa sites, and Intiqa was trained for 2 epochs to match its scores. None of the test questions or their relevant passages were in training. The training data is not released.
License
Apache-2.0, like the base model, mmarco-mMiniLMv2-L12-H384-v1.