gn-trvn-bert-tiny-stage2

SentenceTransformer checkpoint fine-tuned for Vietnamese retrieval.

Evaluation

• Truncate dims: [384, 256, 128, 64]

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
384	0.456316	0.634294	0.701852	0.762315	0.551639	0.573543	0.588873	0.524094	0.533167	0.537396	0.634294	0.524094
256	0.446844	0.624487	0.693582	0.756252	0.5419	0.56431	0.580202	0.514374	0.523662	0.528049	0.624487	0.514374
128	0.425247	0.606829	0.67679	0.742337	0.521721	0.544355	0.560955	0.493421	0.502769	0.507339	0.606829	0.493421
64	0.38085	0.568523	0.643039	0.713811	0.480685	0.504813	0.522732	0.451458	0.461432	0.466364	0.568523	0.451458

Benchmark (x100, GreenNode/MTEB layout):

method	MAP@5	MRR@5	NDCG@5	Recall@5	Mean	HitRate@1	HitRate@5	HitRate@10	HitRate@20
384	52.41	52.41	55.16	63.43	55.85	45.63	63.43	70.19	76.23
256	51.44	51.44	54.19	62.45	54.88	44.68	62.45	69.36	75.63
128	49.34	49.34	52.17	60.68	52.88	42.52	60.68	67.68	74.23
64	45.15	45.15	48.07	56.85	48.8	38.08	56.85	64.3	71.38

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
384	0.28934	0.468274	0.545685	0.614213	0.381232	0.405831	0.422999	0.352517	0.362577	0.367314	0.468274	0.352623
256	0.265228	0.459391	0.532995	0.623096	0.367553	0.391337	0.41414	0.337014	0.346827	0.353204	0.459391	0.337119
128	0.246193	0.420051	0.492386	0.572335	0.336712	0.35977	0.379862	0.309306	0.318628	0.324174	0.420051	0.309095
64	0.218274	0.359137	0.42132	0.483503	0.292999	0.313054	0.328859	0.271468	0.279718	0.284104	0.358503	0.271024

Benchmark (x100, GreenNode/MTEB layout):

method	MAP@5	MRR@5	NDCG@5	Recall@5	Mean	HitRate@1	HitRate@5	HitRate@10	HitRate@20
384	35.26	35.25	38.12	46.83	38.87	28.93	46.83	54.57	61.42
256	33.71	33.7	36.76	45.94	37.53	26.52	45.94	53.3	62.31
128	30.91	30.93	33.67	42.01	34.38	24.62	42.01	49.24	57.23
64	27.1	27.15	29.3	35.85	29.85	21.83	35.91	42.13	48.35

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
384	0.37207	0.628418	0.736328	0.811035	0.33977	0.370704	0.395296	0.469816	0.484128	0.489477	0.357178	0.265946
256	0.356934	0.60791	0.708008	0.795898	0.325689	0.356082	0.381456	0.451318	0.464667	0.470962	0.343506	0.253817
128	0.32959	0.581543	0.688965	0.769531	0.306874	0.337021	0.360692	0.423079	0.437518	0.443241	0.32666	0.238192
64	0.283203	0.505371	0.617676	0.72168	0.264107	0.295047	0.321912	0.36543	0.380457	0.387663	0.282471	0.203589

Benchmark (x100, GreenNode/MTEB layout):

method	MAP@5	MRR@5	NDCG@5	Recall@5	Mean	HitRate@1	HitRate@5	HitRate@10	HitRate@20
384	26.59	46.98	33.98	35.72	35.82	37.21	62.84	73.63	81.1
256	25.38	45.13	32.57	34.35	34.36	35.69	60.79	70.8	79.59
128	23.82	42.31	30.69	32.67	32.37	32.96	58.15	68.9	76.95
64	20.36	36.54	26.41	28.25	27.89	28.32	50.54	61.77	72.17