ViRanker: A BGE-M3 & Blockwise Parallel Transformer Cross-Encoder for Vietnamese Reranking

Kieu-Linh Nguyen; Phuong-Nam Dang; Thanh-Hieu Pham

REVIEW 4 major objections 7 minor 20 references

ViRanker: A BGE-M3 & Blockwise Parallel Transformer Cross-Encoder for Vietnamese Reranking

T0 review · 4 major / 7 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read ViRanker, a Vietnamese cross-encoder reranker built on BGE-M3 with Rotary Position Encoding and Blockwise Parallel Transformer, claims top-3 accuracy on MMARCO-VI that surpasses multilingual baselines and rivals PhoRanker.

desk verdict A plausible and openly released Vietnamese reranker, but its headline numbers rest on an author-hosted benchmark with no external validation—worth refereeing, not yet worth trusting. read the letter →

arxiv 2509.09131 v1 pith:IE67AWL5 submitted 2025-09-11 cs.CL cs.AI

Phuong-Nam Dang , Kieu-Linh Nguyen , Thanh-Hieu Pham This is my paper · ORCID

classification cs.CLcs.AI

keywords Vietnamesererankingcross-encoderBGE-M3BlockwiseParallelTransformerRoPEhard-negativeminingMMARCO-VIlow-resourceNLP

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper is trying to establish that a Vietnamese-specific cross-encoder reranker can close the gap left by multilingual models, especially where users look only at the top few results. ViRanker combines BGE-M3 with Rotary Position Encoding and the Blockwise Parallel Transformer, and is trained on a newly assembled 8 GB Vietnamese corpus with 3.5 million query–document triplets. On the MMARCO-VI test set the authors report NDCG@3=0.6815 and MRR@3=0.6641, outperforming multilingual BGE baselines and slightly edging PhoRanker at the top of the ranking. If those numbers hold, Vietnamese retrieval and search applications gain an open-weight reranker that improves first-page precision without requiring high-end hardware.

What carries the argument

The central object is ViRanker itself: a BGE-M3 encoder repurposed as a cross-encoder with RoPE (a positional scheme that rotates feature pairs to encode both absolute and relative word order) and BPT (a blockwise mechanism that interleaves attention and feed-forward computation, extending practical context length with lower memory use). Around that architecture, the load-bearing training device is hybrid hard-negative mining: each pseudo-query generated by the Inverse Cloze Task gets BM25 candidates, which are reranked by dense similarity and diversified with Maximal Marginal Relevance, and the three hardest non-relevant passages become negatives. A triplet ranking loss then forces the mode

What would settle it

Run ViRanker on a newly built Vietnamese query–passage set—constructed by different annotators, with queries and labels independent of any data used to train or select the model—and compare NDCG@3 and MRR@3 against the BGE baselines. If the early-rank advantage shrinks or reverses, the MMARCO-VI result was benchmark-specific rather than a general Vietnamese reranking improvement.

Watch

Extended reading notes

Core claim

The central claim is that ViRanker gives Vietnamese retrieval a strong early-rank accuracy advantage through a specific architectural combination. ViRanker adapts the BGE-M3 multilingual encoder into a cross-encoder with an MLP scoring head, replaces absolute position encodings with RoPE to better capture Vietnamese word order and diacritics, and substitutes Blockwise Parallel Transformer attention for FlashAttention to handle long documents efficiently. Training uses 3.5 million triplets mined from an 8 GB corpus with a hybrid hard-negative strategy combining BM25 candidates, dense similarity reranking, and MMR diversification. On MMARCO-VI, the model reaches NDCG@3=0.6815 and MRR@3=0.6641,

Load-bearing premise

The load-bearing premise is that the MMARCO-VI test set is a valid, independently constructed benchmark for Vietnamese reranking; the paper cites only a Hugging Face card under the same account that releases ViRanker and gives no details on query construction, relevance labeling, or contamination checks.

Editorial extensions

If this is right

ViRanker gives Vietnamese retrieval pipelines a top-3 precision boost over multilingual BGE rerankers on MMARCO-VI, exactly where user attention concentrates.
PhoRanker keeps a small lead at NDCG@10, so the two Vietnamese rerankers are complementary rather than interchangeable: ViRanker is stronger at early ranks, PhoRanker slightly deeper.
The reported throughput on A100, V100, and T4 means the accuracy gain does not require an expensive deployment stack, making the model usable in practical Vietnamese search systems.
The training recipe—ICT pseudo-queries plus hybrid hard-negative mining—can be reused to build reranking datasets for other low-resource languages.
Open release of the model makes the results independently checkable and gives Vietnamese information retrieval a reproducible baseline.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Because the paper does not report ablations isolating RoPE or BPT, the cleanest test of its architectural story is to retrain with each modification disabled; if early-rank scores stay flat, the gains may come mainly from the larger, cleaner training corpus rather than the new components.
The error analysis suggests a natural hybrid system: route short factoid queries, where ViRanker is weak, to a fact-oriented reranker or retriever, and use ViRanker for descriptive and medium-length queries.
Since the MMARCO-VI benchmark card appears under the same account that releases ViRanker and the paper gives no details on labeling or contamination checks, an independently constructed Vietnamese benchmark is the decisive next test.
The BPT design point implies a generalizable lesson for long-context reranking: blockwise parallel attention may help other low-resource rerankers handle 1024-token inputs at lower memory cost.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The short version: ViRanker is a legitimate engineering contribution—a Vietnamese cross-encoder reranker built from BGE-M3 plus RoPE, BPT, and a custom hard-negative mining pipeline—and its reported early-rank numbers are plausible. But the evaluation rests on a benchmark whose provenance is the same Hugging Face account that publishes the model, and the paper gives no independent evidence for the MMARCO-VI test set. Until that benchmark is externally validated or the dataset construction is documented, the headline NDCG@3/MRR@3 gains should be treated as unverified.

What's genuinely good: the model is openly released, the training corpus is described in concrete detail (8GB, 3.5M chunks, sources enumerated), and the authors are unusually upfront in their limitations section about short factoid queries, long narrative drift, and missing robustness checks. The architecture combination is new for Vietnamese, even if each ingredient is known. That is a real artifact with real user demand, not a toy.

Soft spots, in order of importance. First, the benchmark. The paper cites MMARCO-VI by a Hugging Face dataset card under namdp-ptit, the same account that hosts ViRanker, and gives no details on query labeling, contamination checks, or how it relates to existing Vietnamese IR benchmarks. The stress-test is right to call this load-bearing: every headline number depends on it. Second, the 'averaged across three seeds' claim is unverifiable—no seed values, no variance, no significance test. Third, the PhoBERT baseline at NDCG@3 of 0.0963 is implausibly low; fine-tuning should give it at least random-level lift, so it suggests a misconfiguration. That doesn't invalidate the comparison, but it doesn't help. Fourth, the training pipeline is described but no code or hyperparameter search details are given, so reproducibility is weaker than the 'open release' claim implies. The comparative analysis with PhoRanker is also mostly assertion, not shown data.

Does the central argument hold? The claim that ViRanker is a strong Vietnamese reranker is plausible, and the architecture choices are reasonable. The claim that it systematically beats baselines is not established until the benchmark and variance issues are addressed. This is a solid engineering paper that needs a real referee, not a desk reject, because the artifact is useful and the flaws are specific and fixable.

Who's it for: anyone building Vietnamese RAG or search pipelines, and researchers working on low-resource rerankers. I'd bring it to a reading group only after the benchmark question is resolved. My recommendation: send it to peer review with a request for external benchmark validation or full dataset documentation, and a mandatory variance table.

Referee Report

4 major / 7 minor

Summary. The paper introduces ViRanker, a Vietnamese cross-encoder reranker built on the BGE-M3 encoder with claimed integration of RoPE and the Blockwise Parallel Transformer (BPT). The model is trained on a curated 8 GB Vietnamese corpus using hybrid hard-negative mining, and the authors report strong early-rank accuracy on the MMARCO-VI benchmark: NDCG@3 = 0.6815 and MRR@3 = 0.6641, surpassing multilingual BGE rerankers and slightly outperforming PhoRanker at top cutoffs. The paper also reports inference efficiency on A100, V100, and T4 GPUs and states that the model is released openly on Hugging Face.

Significance. If the reported results are reliable, ViRanker would be a practically useful open-weight reranker for Vietnamese, a language with few dedicated reranking resources. The efficiency measurements in Table 1 are useful for deployment decisions, and the release of the model is a positive step for reproducibility. However, the central accuracy claim depends entirely on the MMARCO-VI benchmark, whose independence and construction are not documented, and the reported numbers lack any variance or significance information. These issues must be addressed before the claimed superiority over baselines can be accepted.

major comments (4)

[§4, Tables 4–5, Ref. [4]] The central empirical claim rests on the MMARCO-VI test set, but this dataset is cited only as a Hugging Face card hosted under the namdp-ptit account, the same account that releases ViRanker. The manuscript provides no details on how queries were constructed, how relevance was labeled, or whether the test set overlaps with the 8 GB training corpus described in §3.2 (Vietnamese Wikipedia, GitHub, books). Without such documentation, the reported NDCG@3 = 0.6815 and MRR@3 = 0.6641 cannot be interpreted as independent evidence of ranking ability. The authors should provide a complete dataset card, contamination checks, and ideally an evaluation on an externally constructed Vietnamese retrieval benchmark.
[§4, Tables 4–5] The text says results are 'averaged across three random seeds,' but no variance, confidence intervals, per-seed results, or significance tests are reported. The gap between ViRanker and PhoRanker at NDCG@3 is 0.019, which may be within noise. Without standard deviations or statistical testing, the claim that ViRanker 'surpasses' PhoRanker at top ranks is not supported. Please report the full distribution or at least standard deviations and perform pairwise significance tests.
[Table 4, PhoBERT baseline row] The fine-tuned PhoBERT reranker scores NDCG@3 = 0.0963 and MRR@3 = 0.0883, which are implausibly low for a fine-tuned reranker on a Vietnamese passage test set. This suggests a misconfiguration (e.g., wrong checkpoint, input truncation, or label mismatch). Since the paper uses this low baseline to argue that 'architectural innovations and careful data curation' are beneficial, the baseline must be validated or removed. Please check the evaluation script and report the exact checkpoint and inference settings.
[§3.3, Figure 1/2 and Table 3] The architectural contribution is not empirically isolated. The paper asserts that RoPE and BPT improve accuracy, but no ablation is provided to separate these modifications from the effect of the 8 GB training corpus and hard-negative sampling. Additionally, BGE-M3 is a pretrained multilingual model that may already use rotary position embeddings; the claim that RoPE 'replaces absolute positional encoding' needs clarification with respect to the actual BGE-M3 backbone. An ablation study (backbone with/without RoPE, with/without BPT) is needed to support the design-choices narrative.

minor comments (7)

[Table 1] Typo in the header: 'quesries/sec' should be 'queries/sec'.
[§3.2 and §2] Figure numbering is inconsistent: §3.2 refers to 'Figure 2' for the hard-negative mining strategy, but Figure 2 in §2 is the context-length plot. Please renumber figures consistently.
[§2, Memory Bank implementation] The PyTorch snippet in Figure 1 computes a contrastive loss with a queue, but the queue is updated with positive embeddings from the current batch and the labels are always `torch.arange(...)`. This is not a standard Memory Bank update (which typically uses a momentum encoder or stored features) and may confuse readers. If this is not the actual training loss used, it should be removed or clarified.
[§3.2] The hard-negative mining description mentions BM25, BGE-M3 embeddings, and MMR but does not specify the diversity threshold or how the 'three most challenging' negatives are selected. Add details for reproducibility.
[§3.4] The hyperparameter table lists 'Memory Bank Size = 512' and 'Batch Size = 512' but the memory-bank mechanism from §2 is not mentioned in the training procedure. Clarify whether a memory bank is used and how it interacts with the triplet loss described in §3.3.
[References [3], [4]] PhoRanker and MMARCO-VI are cited only via Hugging Face model/dataset cards, not via peer-reviewed descriptions. If no published reference exists, state this explicitly and provide the exact URLs and versioning information.
[Introduction] The claim of 'approximately 5,000 monthly users, peaking at over 23,000' is not a scientific evaluation metric and cannot be verified from the paper. It may be omitted or moved to a non-empirical section.

Circularity Check

1 steps flagged · score 4.0 of 10

Headline accuracy numbers rest on an author-hosted benchmark; no direct derivation circularity but a closed evaluation loop.

self citation load bearing [Section 4 (Tables 4–5) and References [4]; model release link in Section 1]
"Evaluated on the MMARCO-VI benchmark, ViRanker achieves strong early-rank accuracy (NDCG@3 = 0.6815, MRR@3 = 0.6641), surpassing multilingual baselines and competing closely with PhoRanker. ... [4] Nguyen, V.H., Tran, T.T., et al.: MMARCO-VI: A Vietnamese benchmark for passage retrieval and reranking. Hugging Face dataset card (2023). Available at: https://huggingface.co/datasets/namdp-ptit/MMARCO-VI."

The only benchmark supporting the paper's headline claims is hosted at namdp-ptit/MMARCO-VI, the same Hugging Face account that publishes namdp-ptit/ViRanker. The paper gives no query-construction, labeling, or contamination details for the benchmark and cites only this self-hosted dataset card. Thus the numeric superiority (NDCG@3=0.6815, MRR@3=0.6641) is evaluated on a dataset released by the same authors, making the evaluation loop closed: the claimed result is not anchored to any independent, externally hosted test set. This is load-bearing because every comparative claim in the abstract and Section 4 depends on these numbers.

full rationale

The paper does not contain a formal derivation that reduces to its own inputs; the training pipeline and architecture are described as independent contributions. However, the central claim of outperformance is entirely supported by a benchmark hosted on the same Hugging Face account as the model, with no external validation, documentation of how queries were constructed, relevance labels, or contamination checks. This is a self-referential evaluation loop rather than a formal circular derivation. Because the architectural and training components have independent content but the headline evaluation is not externally anchored, a moderate circularity score of 4 is assigned. If the dataset were independently vetted, externally hosted, or code-reproduced, the score would drop to 1.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

No new physical or abstract entities are introduced. The model's behavior rests on standard pretrained components, but the evaluation loop is partly self-referential and key training choices are hand-selected without ablation.

free parameters (7)

Learning rate = 5e-5
Hand-tuned value in Table 3; no search or sensitivity analysis.
Batch size = 512
Chosen for GPU memory and stability; no analysis of effect on NDCG.
Max sequence length = 1024 tokens
Data chunks joined to 1024; score may depend on this truncation.
Number of hard negatives per query = 3
Selected from top 20 BM25 candidates after MMR; no ablation on negative count.
BPT chunk size = 32
Table 3 lists Query/Key chunk 32; no sensitivity analysis.
Memory Bank Size = 512
Listed in Table 3 although training uses triplet loss; role unclear.
Training epochs = 10
No early stopping or epoch search reported.

assumptions (5)

domain assumption MMARCO-VI is a valid, independently constructed evaluation benchmark.
Tables 4 and 5 rely entirely on this test set; ref [4] points to an author-hosted HF dataset rather than an external benchmark.
domain assumption BGE-M3 checkpoints are suitable initialization for a Vietnamese cross-encoder.
Used as backbone in Section 3.3 without comparative pretraining study.
domain assumption Blockwise Parallel Transformer can be dropped into an encoder cross-encoder without changing ranking behavior.
Section 3.3 replaces FlashAttention with BPT, but no architectural analysis or standalone experiment verifies this.
domain assumption LLM-based spelling correction improves training data quality.
Section 3.2 applies Gemma-1.5 and GPT-4o-mini corrections; no validation that corrections do not introduce errors or domain drift.
domain assumption Triplet loss with hard negatives optimizes NDCG/MRR.
Section 3.3 adopts triplet loss [18]; the link to ranking metrics is not formally established.

how reviews work

0 comments

Cite this review

Pith. "Pith review of ViRanker: A BGE-M3 & Blockwise Parallel Transformer Cross-Encoder for Vietnamese Reranking." pith.science (2026). https://pith.science/paper/IE67AWL5

@misc{pith2026250909131,
  author       = {Pith},
  title        = {Pith review of: ViRanker: A BGE-M3 & Blockwise Parallel Transformer Cross-Encoder for Vietnamese Reranking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IE67AWL5}},
  note         = {Machine review of arXiv:2509.09131}
}

read the original abstract

This paper presents ViRanker, a cross-encoder reranking model tailored to the Vietnamese language. Built on the BGE-M3 encoder and enhanced with the Blockwise Parallel Transformer, ViRanker addresses the lack of competitive rerankers for Vietnamese, a low-resource language with complex syntax and diacritics. The model was trained on an 8 GB curated corpus and fine-tuned with hybrid hard-negative sampling to strengthen robustness. Evaluated on the MMARCO-VI benchmark, ViRanker achieves strong early-rank accuracy, surpassing multilingual baselines and competing closely with PhoRanker. By releasing the model openly on Hugging Face, we aim to support reproducibility and encourage wider adoption in real-world retrieval systems. Beyond Vietnamese, this study illustrates how careful architectural adaptation and data curation can advance reranking in other underrepresented languages.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 5 linked inside Pith

[1]

In: Advances in Neural Information Processing Systems (NeurIPS), 2020

Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Fan, A., Barta, M., Khandelwal, U., Lewis, M.: Retrieval -Augmented Generation for Knowledge -Intensive NLP Tasks. In: Advances in Neural Information Processing Systems (NeurIPS), 2020

2020
[2]

In: Findings of EMNLP, pp

Nguyen, D.Q., Nguyen, A.T.: PhoBERT: Pre -trained language models for Vietnamese. In: Findings of EMNLP, pp. 1037–1042 (2020)

2020
[3]

Hugging Face model card (2023)

It, D.A.B.: PhoRanker: Vietnamese cross -encoder reranker. Hugging Face model card (2023). Available at: https://huggingface.co/itdainb/PhoRanker. 9

2023
[4]

Hugging Face dataset card (2023)

Nguyen, V.H., Tran, T.T., et al.: MMARCO -VI: A Vietnamese benchmark for passage re- trieval and reranking. Hugging Face dataset card (2023). Available at: https://hugging- face.co/datasets/namdp-ptit/MMARCO-VI

2023
[5]

In: Advances in Neural Information Processing Systems (NeurIPS), 2022

Dao, T., Fu, D., Ermon, S., Rudra, A., Ré, C.: FlashAttention: Fast and memory -efficient exact attention with IO-awareness. In: Advances in Neural Information Processing Systems (NeurIPS), 2022

2022
[6]

arXiv preprint arXiv:2402.03216 (2024)

Chen, J., Xiao, S., Zhang, P., Luo, K., Lian, D.: BGE M3-Embedding: Multi-lingual, Multi- functionality, Multi -granularity Text Embeddings through Self -Knowledge Distillation. arXiv preprint arXiv:2402.03216 (2024). https://arxiv.org/abs/2402.03216

arXiv 2024
[7]

arXiv preprint arXiv:2104.09864 (2021)

Su, J., Lu, Y., Pan, S., Wen, B., Liu, Y.: RoFormer: Enhanced Transformer with Rotary Position Embedding. arXiv preprint arXiv:2104.09864 (2021)

arXiv 2021
[8]

In: Proceedings of the International Conference on Machine Learning (ICML), 2024

Zhang, H., Li, Q., Zhao, T.: Blockwise Parallel Transformers for Efficient Long -Context Processing. In: Proceedings of the International Conference on Machine Learning (ICML), 2024

2024

Show all 20 references

[9]

Foundations and Trends in Information Retrieval 3(4), 333–389 (2009)

Robertson, S., Zaragoza, H.: The probabilistic relevance framework: BM25 and beyond. Foundations and Trends in Information Retrieval 3(4), 333–389 (2009)

2009
[10]

In: Proceedings of the 21st Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pp

Carbonell, J., Goldstein, J.: The Use of MMR, Diversity -Based Reranking for Reordering Documents and Producing Summaries. In: Proceedings of the 21st Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pp. 335– 336 (1998). https://d...

1998
[11]

arXiv preprint arXiv:1901.04085 (2019)

Nogueira, R., Cho, K.: Passage re -ranking with BERT. arXiv preprint arXiv:1901.04085 (2019)

1901 arXiv
[12]

In: Proceeding s of NAACL, pp

Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In: Proceeding s of NAACL, pp. 4171 –4186 (2019)

2019
[13]

In: Proceedings of ACL, pp

Conneau, A., Khandelwal, K., Goyal, N., Chaudhary, V., Wenzek, G., Guzmán, F., Grave, E., Ott, M., Zettlemoyer, L., Stoyanov, V.: Unsupervised Cross -lingual Representation Learning at Scale. In: Proceedings of ACL, pp. 8440–8451 (2020)

2020
[14]

In: Advances in Neural Information Processing Systems (Neu- rIPS), 2017

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Polosukhin, I.: Attention is all you need. In: Advances in Neural Information Processing Systems (Neu- rIPS), 2017

2017
[15]

In: Proceedings of NAACL, 2018

Shaw, P., Uszkoreit, J., Vaswani, A.: Self -attention with relative position representations. In: Proceedings of NAACL, 2018

2018
[16]

Hugging Face model card (2024)

BAAI.: Gemma -Reranker: Large -scale multilingual reranker. Hugging Face model card (2024). Available at: https://huggingface.co/BAAI/bge-reranker-v2-gemma

2024
[17]

Hugging Face model card (2023)

BAAI.: BGE-Reranker-V2-M3: Multilingual passage reranker. Hugging Face model card (2023). Available at: https://huggingface.co/BAAI/bge-reranker-v2-m3

2023
[18]

In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp

Schroff, F., Kalenichenko, D., Philbin, J.: FaceNet: A Unified Embedding for Face Recog- nition and Cluster ing. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 815–823 (2015). https://arxiv.org/abs/1503.03832

2015 arXiv
[19]

In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020

He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum Contrast for Unsupervised Visual Representation Learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020

2020
[20]

arXiv preprint arXiv:2501.01234 (2025)

Liu, Y., Xu, W., Huang, K., et al.: A survey on long-context language models. arXiv preprint arXiv:2501.01234 (2025)

2025 arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

In: Advances in Neural Information Processing Systems (NeurIPS), 2020

Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Fan, A., Barta, M., Khandelwal, U., Lewis, M.: Retrieval -Augmented Generation for Knowledge -Intensive NLP Tasks. In: Advances in Neural Information Processing Systems (NeurIPS), 2020

2020

[2] [2]

In: Findings of EMNLP, pp

Nguyen, D.Q., Nguyen, A.T.: PhoBERT: Pre -trained language models for Vietnamese. In: Findings of EMNLP, pp. 1037–1042 (2020)

2020

[3] [3]

Hugging Face model card (2023)

It, D.A.B.: PhoRanker: Vietnamese cross -encoder reranker. Hugging Face model card (2023). Available at: https://huggingface.co/itdainb/PhoRanker. 9

2023

[4] [4]

Hugging Face dataset card (2023)

Nguyen, V.H., Tran, T.T., et al.: MMARCO -VI: A Vietnamese benchmark for passage re- trieval and reranking. Hugging Face dataset card (2023). Available at: https://hugging- face.co/datasets/namdp-ptit/MMARCO-VI

2023

[5] [5]

In: Advances in Neural Information Processing Systems (NeurIPS), 2022

Dao, T., Fu, D., Ermon, S., Rudra, A., Ré, C.: FlashAttention: Fast and memory -efficient exact attention with IO-awareness. In: Advances in Neural Information Processing Systems (NeurIPS), 2022

2022

[6] [6]

arXiv preprint arXiv:2402.03216 (2024)

Chen, J., Xiao, S., Zhang, P., Luo, K., Lian, D.: BGE M3-Embedding: Multi-lingual, Multi- functionality, Multi -granularity Text Embeddings through Self -Knowledge Distillation. arXiv preprint arXiv:2402.03216 (2024). https://arxiv.org/abs/2402.03216

arXiv 2024

[7] [7]

arXiv preprint arXiv:2104.09864 (2021)

Su, J., Lu, Y., Pan, S., Wen, B., Liu, Y.: RoFormer: Enhanced Transformer with Rotary Position Embedding. arXiv preprint arXiv:2104.09864 (2021)

arXiv 2021

[8] [8]

In: Proceedings of the International Conference on Machine Learning (ICML), 2024

Zhang, H., Li, Q., Zhao, T.: Blockwise Parallel Transformers for Efficient Long -Context Processing. In: Proceedings of the International Conference on Machine Learning (ICML), 2024

2024