bert-base-stage2-sbert-v3

SentenceTransformer checkpoint fine-tuned for Vietnamese legal retrieval.

Evaluation

• Truncate dims: [768, 512, 256, 128]

UIT-ViQuAD2.0

method	Accuracy@1	Accuracy@3	Accuracy@5	Accuracy@10	NDCG@3	NDCG@5	NDCG@10	MRR@3	MRR@5	MRR@10	Recall@5	Recall@10	Recall@100	MAP@100
BM25	0.606492	0.75565	0.805643	0.861663	0.694646	0.715313	0.733532	0.673492	0.685004	0.692583	0.805643	0.861663	0.966717	0.69741
768	0.491576	0.675661	0.743186	0.819477	0.599579	0.627532	0.652369	0.573255	0.588849	0.599196	0.743186	0.819477	0.965347	0.605738
512	0.483632	0.667032	0.733872	0.811533	0.591382	0.618992	0.644249	0.565197	0.580564	0.59107	0.733872	0.811533	0.962745	0.597985
256	0.461992	0.653061	0.718258	0.796466	0.573756	0.600622	0.626045	0.546341	0.561256	0.571828	0.718258	0.796466	0.957677	0.579092
128	0.433639	0.61471	0.684427	0.768936	0.53931	0.567997	0.595255	0.513263	0.529165	0.540371	0.684427	0.768936	0.949185	0.548269

Zalo-Legal

method	Accuracy@1	Accuracy@3	Accuracy@5	Accuracy@10	NDCG@3	NDCG@5	NDCG@10	MRR@3	MRR@5	MRR@10	Recall@5	Recall@10	Recall@100	MAP@100
BM25	0.379442	0.616751	0.69797	0.77665	0.515637	0.549114	0.57497	0.48181	0.500402	0.511188	0.696701	0.775381	0.923858	0.517283
768	0.323604	0.531726	0.59264	0.676396	0.445228	0.470463	0.497713	0.416667	0.430372	0.441706	0.591371	0.675127	0.926396	0.450915
512	0.309645	0.544416	0.591371	0.67132	0.446587	0.46611	0.492081	0.414129	0.424725	0.435511	0.590102	0.670051	0.918147	0.444509
256	0.305838	0.519036	0.591371	0.663706	0.431393	0.461401	0.484796	0.402284	0.418718	0.428379	0.590102	0.662437	0.913071	0.43801
128	0.279188	0.496193	0.562183	0.651015	0.404668	0.432222	0.460904	0.374154	0.389319	0.401135	0.560914	0.649746	0.899112	0.410837