gn-trvn-bamibert

SentenceTransformer checkpoint fine-tuned for Vietnamese retrieval.

Evaluation

• Truncate dims: [768, 512, 256, 128]

GN-TRVN

method	Accuracy@1	Accuracy@5	Accuracy@10	Accuracy@20	NDCG@5	NDCG@10	NDCG@20	MRR@5	MRR@10	MRR@20	Recall@5	MAP@5
768	0.481685	0.661256	0.724009	0.783884	0.577989	0.598381	0.613576	0.550237	0.558709	0.56291	0.661256	0.550237
512	0.477494	0.657232	0.720963	0.779973	0.573751	0.594448	0.609431	0.54593	0.554524	0.55867	0.657232	0.54593
256	0.466095	0.646168	0.712134	0.772177	0.562587	0.584017	0.599228	0.534716	0.543616	0.547807	0.646168	0.534716
128	0.447794	0.630857	0.697969	0.762147	0.545557	0.567357	0.583642	0.517149	0.526202	0.530703	0.630857	0.517149

Benchmark (x100, GreenNode/MTEB layout):

method	MAP@5	MRR@5	NDCG@5	Recall@5	Mean	HitRate@1	HitRate@5	HitRate@10	HitRate@20
768	55.02	55.02	57.8	66.13	58.49	48.17	66.13	72.4	78.39
512	54.59	54.59	57.38	65.72	58.07	47.75	65.72	72.1	78
256	53.47	53.47	56.26	64.62	56.95	46.61	64.62	71.21	77.22
128	51.71	51.71	54.56	63.09	55.27	44.78	63.09	69.8	76.21

Zalo-Legal

method	Accuracy@1	Accuracy@5	Accuracy@10	Accuracy@20	NDCG@5	NDCG@10	NDCG@20	MRR@5	MRR@10	MRR@20	Recall@5	MAP@5
768	0.326142	0.559645	0.652284	0.73731	0.447889	0.477922	0.499286	0.411379	0.423554	0.429348	0.55901	0.41084
512	0.326142	0.560914	0.645939	0.724619	0.449285	0.476837	0.496951	0.412775	0.423913	0.429544	0.560279	0.412236
256	0.321066	0.535533	0.640863	0.727157	0.433658	0.468229	0.490136	0.400402	0.414691	0.420754	0.534898	0.399894
128	0.305838	0.521574	0.615482	0.703046	0.416495	0.44708	0.469304	0.382382	0.394858	0.400999	0.520939	0.381874

Benchmark (x100, GreenNode/MTEB layout):

method	MAP@5	MRR@5	NDCG@5	Recall@5	Mean	HitRate@1	HitRate@5	HitRate@10	HitRate@20
768	41.08	41.14	44.79	55.9	45.73	32.61	55.96	65.23	73.73
512	41.22	41.28	44.93	56.03	45.86	32.61	56.09	64.59	72.46
256	39.99	40.04	43.37	53.49	44.22	32.11	53.55	64.09	72.72
128	38.19	38.24	41.65	52.09	42.54	30.58	52.16	61.55	70.3

VieQuADRetrieval

method	Accuracy@1	Accuracy@5	Accuracy@10	Accuracy@20	NDCG@5	NDCG@10	NDCG@20	MRR@5	MRR@10	MRR@20	Recall@5	MAP@5
768	0.428223	0.696777	0.786133	0.85498	0.379656	0.408516	0.429394	0.531616	0.543734	0.548618	0.393555	0.297538
512	0.419922	0.69043	0.782227	0.84668	0.373495	0.403479	0.423687	0.524154	0.536449	0.540945	0.387939	0.291659
256	0.408691	0.668945	0.749023	0.82373	0.361856	0.38761	0.410144	0.509351	0.520151	0.525333	0.374756	0.282524
128	0.367676	0.637695	0.720703	0.806641	0.337365	0.363332	0.386991	0.47002	0.481252	0.487374	0.356445	0.261222

Benchmark (x100, GreenNode/MTEB layout):

method	MAP@5	MRR@5	NDCG@5	Recall@5	Mean	HitRate@1	HitRate@5	HitRate@10	HitRate@20
768	29.75	53.16	37.97	39.36	40.06	42.82	69.68	78.61	85.5
512	29.17	52.42	37.35	38.79	39.43	41.99	69.04	78.22	84.67
256	28.25	50.94	36.19	37.48	38.21	40.87	66.89	74.9	82.37
128	26.12	47	33.74	35.64	35.63	36.77	63.77	72.07	80.66