bge-reranker-dienbien
Cross-encoder reranker fine-tuned (LoRA, merged) from BAAI/bge-reranker-v2-m3
on Vietnamese legal/administrative documents (Điện Biên province public
resolutions and decisions) for the VAIC2026 Paperless Meetings project.
It is stage 2 of a two-stage retrieval pipeline: multilingual-e5-large fetches
candidates by embedding cosine, then this model re-scores each (question, passage)
pair so the passage that actually answers the question ranks first — which is what
makes page/article citations correct.
Held-out results (18 clean questions, none seen in training)
| Setup | Recall@1 | MRR | nDCG@5 |
|---|---|---|---|
| e5 retrieval only | 0.722 | 0.819 | 0.865 |
| + bge-reranker-v2-m3 (base) | 0.833 | 0.898 | 0.924 |
| + this fine-tune | 0.889 | 0.944 | 0.959 |
Usage
from sentence_transformers import CrossEncoder
model = CrossEncoder("jakethepinkpanther/bge-reranker-dienbien")
scores = model.predict([("Điều kiện hưởng trợ cấp là gì?", "Điều 5. ...")])