README.md

---
library_name: transformers
pipeline_tag: text-ranking
license: mit
language:
- en
- zh
base_model:
- jhu-clsp/mmBERT-base
tags:
- reranker
- modernbert
- English
- zh-tw
- zh-cn
---

# AuroraX: A Fast Cross-Lingual Reranker Bridging English and Chinese

AuroraX is a lightweight yet powerful cross-lingual reranker built upon the [mmBERT-base](https://huggingface.co/jhu-clsp/mmBERT-base) architecture.
It is designed to bridge **Traditional Chinese**, **Simplified Chinese** and **English**, enabling high-quality semantic ranking across languages with remarkable efficiency.

Despite having only 110M non-embedding parameters, AuroraX achieves comparable performance to state-of-the-art rerankers that are twice as large.
Its design emphasizes both speed and language adaptability, making it ideal for real-world multilingual retrieval and re-ranking applications.

✨ Key Features:
- 🌏 **Cross-Lingual Understanding** — Trained to handle English, Traditional Chinese, and Simplified Chinese seamlessly.  
- ⚡ **Lightweight & Fast** — Only 110M parameters (non-embedding), optimized for latency-sensitive pipelines.  
- 🎯 **SOTA-Level Accuracy** — Comparable or superior to larger rerankers on Chinese and English benchmarks.

---

## Evaluation

### Monolingual Benchmarks

| Model | Metric | CMedQAv2-reranking (ZH) | T2Reranking (ZH) | **ZH AVG** | AskUbuntuDupQuestions (EN) | HUMENews21InstructionReranking (EN) | HUMEWikipediaRerankingMultilingual (EN) | SciDocsRR (EN) | **EN AVG** | **Total AVG** |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| **AuroraX-Reranker-Base-v1.0**<br>*(Ours, 300M with 100M non-embed params)* | **mrr@10** | **0.8201** | **0.8554** | **0.8378** | **0.7936** | **1.0000** | **0.9778** | **0.9305** | **0.9255** | **0.8962** |
|  | **mrr@5** | 0.8145 | 0.8514 | 0.8329 | 0.7841 | 1.0000 | 0.9778 | 0.9289 | 0.9227 | 0.8928 |
| **bge-reranker-v2-m3**<br>*(600M params)* | **mrr@10** | 0.8598 | 0.8004 | 0.8301 | 0.7635 | 0.9839 | 0.8750 | 0.9211 | 0.8859 | 0.8673 |
|  | **mrr@5** | 0.8569 | 0.7954 | 0.8262 | 0.7532 | 0.9839 | 0.8750 | 0.9191 | 0.8828 | 0.8639 |
| **jina-reranker-v2-base-multilingual**<br>*(300M params)* | **mrr@10** | 0.2828 | 0.7577 | 0.5203 | 0.7420 | 1.0000 | 0.8761 | 0.9478 | 0.8915 | 0.7677 |
|  | **mrr@5** | 0.2759 | 0.7512 | 0.5136 | 0.7299 | 1.0000 | 0.8761 | 0.9467 | 0.8882 | 0.7633 |

---

### Cross-Lingual (ZH ↔ EN) Results

| Model | inhouse-en2zh (HitRate@5) | inhouse-zh2en (HitRate@5) |
| --- | --- | --- |
| **AuroraX-Reranker-Base-v1.0 (Ours, 300M with 100M non-embed params)** | **0.8459** | **0.9427** |
| **bge-reranker-v2-m3 (600M params)** | 0.8179 | 0.9160 |
| **jina-reranker-v2-base-multilingual (300M params)** | 0.7815 | 0.8855 |

---
## Usage

### Sentence-Transformers

```py
from sentence_transformers import CrossEncoder

model = CrossEncoder("aqweteddy/AuroraX-Reranker-Base-v1.0")
score = model.predict([("What is Deep Learning?", "Deep learning is a subfield of ML...")])
print(score)
```

### Text-Embedding-Inference (API)

1. Install and launch the router:

```bash
text-embeddings-router --model-id aqweteddy/AuroraX-Reranker-Base-v1.0
```

2. Run via REST API:

```bash
curl 127.0.0.1:8080/rerank \
  -X POST \
  -d '{"query": "What is Deep Learning?", "texts": ["Deep Learning is not...", "Deep learning is..."]}' \
  -H 'Content-Type: application/json'
```

---

## Citation

```
@misc{aurorax2025,
  title         = {AuroraX: A Fast Cross-Lingual Reranker Bridging English and Chinese},
  author        = {aqweteddy},
  year          = {2025},
  howpublished  = {\url{https://huggingface.co/aqweteddy/AuroraX-Reranker-Base-v1.0}},
  note          = {Lightweight and powerful eranker for English, Traditional Chinese, and Simplified Chinese}
}
```