Pith. sign in

REVIEW 4 major objections 4 minor 18 references

GRAF: Graph Retrieval Augmented by Facts for Romanian Legal Multi-Choice Question Answering

T0 review · 4 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read A graph-retrieval model, GRAF, beats LLM baselines on Romanian legal multiple-choice exams by grounding each answer in a knowledge graph.

desk verdict Genuinely useful new resources for Romanian legal NLP, but the headline performance claim is a per-branch best-of-many artifact and should not be taken as evidence that GRAF beats SOTA. read the letter →

arxiv 2412.04119 v3 pith:PXC6DHYO submitted 2024-12-05 cs.CL

Cristian-George Crăciun , Răzvan-Alexandru Smădu , Dumitru-Clementin Cercel , Mihaela-Claudia Cercel This is my paper · ORCID
classification cs.CL
keywords Romanianlegalquestionansweringmultiple-choiceQAknowledgegraphretrieval-augmentedgenerationattentionnetworklow-resourceNLPdatasetinformationretrieval
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

This paper claims that a retrieval-augmented multiple-choice QA system for Romanian law can beat strong retrieval and fine-tuned baselines by grounding answers in a knowledge graph. To test this, the authors build and release the first open Romanian legal exam dataset (JuRO), a curated law corpus (CROL), and a legal knowledge graph (Law-RoG). Their method, GRAF, turns each question-choice pair into a small claim graph, retrieves the most relevant law-graph neighbourhood, and aligns the two before scoring. On promotion exams GRAF reaches 60.09% average accuracy against 55.75% for the strongest baseline and leads in six of nine legal branches; on entrance and bar exams it also improves over encoder and LLM baselines. The wider point is that structured legal knowledge can be a viable alternative to larger language models in low-resource settings.

What carries the argument

The load-bearing mechanism is the GRAF pipeline built around Law-RoG, a knowledge graph of 160k nodes and 320k edges extracted from the CROL law corpus. For each candidate choice, an LLM produces a claim graph of (entity; relation; entity) triplets; a BM25-based sampler retrieves the top entities and their one-hop neighbourhoods from Law-RoG; and a Graph Attention Network with separate node and edge transformations encodes both graphs. The paper's central identity is the relevance alignment between encoded claim nodes and the sampled knowledge-graph relations --- cosine similarity $R_{ij} = \cos(h_i^c, h_j)$ --- which carries the extracted facts into the attention head that scores the choice.

What would settle it

Rerun the entrance and bar exam evaluations with the number of correct choices withheld from GRAF at inference, or included in the LLM prompt, and separate scores by whether the question has one or two correct answers. If GRAF's margin over the best LLM baseline disappears or reverses when the answer count is not supplied, then the headline advantage comes from that information rather than from graph retrieval.

Watch

Extended reading notes

Core claim

GRAF's central claim is that claim-aware graph retrieval improves multiple-choice legal QA: instead of retrieving raw legal text, the model first asks an LLM to extract entity-relation-entity claims from the question and each candidate answer, samples a domain-specific subgraph from Law-RoG using BM25 and breadth-first search, then encodes nodes and edges with a relational Graph Attention Network. A cosine-similarity alignment selects the knowledge-graph relations most relevant to the claims, and self-attention combines this evidence with the question-choice encoding to produce a score. On the promotion-exam portion of JuRO, this reaches 60.09% accuracy versus 55.75% for the best LLM baseline, and the paper reports gains in six of nine legal branches, with the largest margins on civil and commercial law. The paper also reports that removing either the claim graph or the knowledge graph drops accuracy, and that simple summation of graph embeddings damages performance further.

Load-bearing premise

The load-bearing premise is that the entrance and bar exam comparisons are fair even though GRAF and the other encoder models are told exactly how many choices are correct while the LLM baselines have to infer that number from the prompt; if that comparison is unfair, the claimed advantage on multi-answer exams is not established.

Editorial extensions

If this is right

  • If GRAF's results hold, Romanian legal MCQA now has an open benchmark and a strong baseline that future methods must match.
  • Legal-domain pre-training is the best backbone choice, since GRAF's largest gains come from jurBERT rather than general-domain Romanian encoders.
  • Both the claim graph and the knowledge graph are load-bearing: ablating either one costs about two accuracy points, and losing both plus collapsing embeddings costs 6.10 points.
  • The released corpus and graph let other researchers do information retrieval and knowledge-graph work on Romanian law without re-scraping official portals.
  • GRAF's performance improves with stronger backbone encoders, so the approach should continue to improve as better Romanian language models appear.

Reading between the lines

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

  • An implication the paper leaves implicit: the claim-graph plus KG-alignment recipe could transfer to other low-resource legal systems, since the knowledge graph itself is built with an LLM rather than hand-crafted resources.
  • A testable extension is to withhold the correct-answer count from GRAF at inference, or give that count to the LLM baselines, to isolate whether the multi-answer advantage comes from graph retrieval or from the evaluation protocol.
  • The paper's LLM-based triplet extraction was only spot-checked by human experts, so a natural stress test is measuring how GRAF behaves when Law-RoG is built from a different seed model or deliberately corrupted.
  • The same alignment machinery could be applied to other fact-verification tasks where a candidate claim must be checked against a structured repository, not just to legal exam questions.
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.

Referee Report

4 major / 4 minor

Summary. The paper introduces three Romanian legal NLP resources and a retrieval-augmented MCQA method: JuRO, a public dataset of 10,836 multiple-choice exam questions; CROL, a corpus of Romanian laws; and Law-RoG, a knowledge graph extracted from CROL. The proposed method, GRAF, builds a claim graph from a question-choice pair with an LLM, retrieves a subgraph from Law-RoG, encodes both with a GAT over a language-model backbone, and scores choices with self-attention. The authors report accuracy on promotion, entrance, and bar exams, comparing GRAF with BERT-family encoders, LLM zero-shot baselines, RAG, and LoRA fine-tuning, and include ablations and topic-level analyses.

Significance. If the empirical claims held, the paper would be a useful contribution to low-resource legal NLP: it is the first open Romanian legal MCQA dataset, provides a structured legal corpus and a Romanian legal knowledge graph, and releases code and resources. The method is described in sufficient detail to reproduce, and the ablation isolates the contribution of the claim graph and the KG. However, the headline claim of exceeding state-of-the-art methods is not currently established because of evaluation asymmetries and reporting choices; the resource contribution is real and likely valuable, but the comparative evaluation needs repair.

major comments (4)
  1. [Appendix D, Eq. (16)] Encoder-based models and GRAF are given the exact number of correct choices |T_i| at inference time because they select the top |T_i| scores, while the LLM baselines are only instructed that 'at most 2 answers are correct' (Appendices I/J). On the entrance and bar exams in Tables 5 and 6, where |T_i| is either 1 or 2, this is an information asymmetry: knowing the exact answer count is a substantial extra signal for multi-answer scoring. The superiority of GRAF in these settings is therefore not established unless all systems are evaluated with the same answer-count information, for example by learning to predict the count or by using a protocol that does not rely on |T_i| at inference.
  2. [Tables 4-6 and Appendix H] The GRAF rows in the main tables are per-branch best-of-many mosaics, not a single fixed model. In Table 4, the 60.09% promotion average combines different backbone models in different branches, e.g., Civil 49.46 is GRAF+jurBERT-base, Penal 52.70 is GRAF+RoBERT-large, and Civil Procedure 78.46 is GRAF+jurBERT-base; the fixed GRAF+jurBERT-large configuration averages 55.81% in Table 12. The LLM LFT baseline row similarly mixes configurations: the best single fixed baseline, FLAN-T5 XXL LFT, averages 55.31%. The fixed-configuration gap on promotion exams is therefore about 0.50 points rather than the 4.34 points implied by comparing the mosaic rows, and the statement in Section 5.3 that GRAF outperforms baselines in 6 of 9 branches is a property of the per-branch best selection. Please report fixed-configuration comparisons and clearly separate them from any per-branch upper-bound analysis.
  3. [Appendix E] The appendix states 'We report the results of a single run.' After aligning configurations, the central advantage on promotion exams is about 0.50 points, and the test sets are small (approximately 50-400 examples per branch in Table 11). Without multiple seeds or paired statistical tests, the reported differences cannot be distinguished from noise. The paper should provide variance estimates across seeds or significance testing, and the abstract and conclusions should be worded to match the strength of the evidence.
  4. [Section 5.4, Table 8] The ablation reports GRAF at 55.61% on promotion exams, but no backbone or configuration is named, and this value matches neither the mosaic row of 60.09% in Table 4 nor any fixed configuration in Table 12 (e.g., GRAF+jurBERT-large at 55.81%, GRAF+RoBERT-large at 55.13%). Since this ablation is the main evidence that both the claim graph and the KG contribute, it must be run with the same configuration used in the main comparison and reported with that configuration; otherwise the effect sizes cannot be interpreted.
minor comments (4)
  1. [Abstract vs. Section 3.2] The abstract says CROL contains 93 documents with modifications from 763 time spans, while Section 3.2 states 768 different versions; please make these numbers consistent.
  2. [Section 4.2] The text refers to 'Algorithm 1' several times, but no algorithm block appears in the manuscript; either include the pseudocode or remove the references.
  3. [Appendix J] The translated FLAN-T5 RAG prompt contains the typo 'documets' instead of 'documents'.
  4. [Appendix F] The appendix contains the typo 'arameter fine-tuning' in the LLM LFT paragraph; it should read 'parameter fine-tuning'.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; minor same-group citations are not load-bearing and the benchmark derivation is independent of GRAF's construction.

full rationale

I walked the claimed derivation chain. GRAF's predictions are produced by a trained encoder that scores each (question, choice) pair, combining a claim graph extracted by Mixtral, a BM25-sampled subgraph from Law-RoG, a GAT encoder, and a learned self-attention scorer. None of these quantities is defined in terms of the target labels, and no parameter is fitted to the reported accuracy. The KG and the RAG baseline both draw on CROL, but a shared retrieval corpus is standard practice and does not force the answer. The same LLM family is used for KG construction and claim extraction, but that is a component choice, not a derivation of the answer. Same-group citations (Smădu et al. 2022; Dima et al. 2024) appear only in related-work positioning, not as evidence for the central claim. The only protocol concern—Appendix D, Eq. 16 hands encoder models the true number of correct choices, |T_i|, while LLM baselines must infer it from the prompt—is an evaluation asymmetry and a form of label leakage, but it is not circularity: the predicted set is still the top-K of learned scores, not the label set itself, so the answer is not constructed from the labels. The best-of-many mosaic in Tables 4–6 weakens the 'exceeds SOTA in most settings' headline, but that is a statistical-reporting and significance issue, not a reduction of the result to its inputs. No uniqueness theorem or self-citation chain forces the architectural choice. Overall, the central derivation is empirically self-contained against the JuRO benchmark, and no circular step meets the quoted-equivalence bar.

Assumptions & free parameters 7 free parameters · 4 assumptions · 1 invented entities

The central claim rests on the hyperparameters of GRAF and on the quality of the automatically built KG and the dataset labels. No physical constants or new theoretical entities are introduced; the only invented resource is Law-RoG, whose accuracy is spot-checked on 10 documents per legal branch.

free parameters (7)
  • BM25 top-k entities = 10
    Subgraph retrieval keeps the top 10 KG entities after BM25 scoring.
  • BFS depth = 1
    Selects neighbors one hop from retrieved entities; larger depths were not explored.
  • Max entities in subgraph = 50
    Caps the sampled graph size for the GAT encoder.
  • GAT attention heads = 6
    Chosen empirically as a balance between non-zero entries and compute (Section G).
  • Learning rate = 1e-7
    Used for all fine-tuning; not tuned per model.
  • Training epochs = 50
    All models trained for 50 epochs with early stopping via best validation checkpoint.
  • LoRA rank = 64
    Used for LLM fine-tuning baseline; not part of GRAF.
assumptions (4)
  • domain assumption Mixtral-8x7B-Instruct extracts accurate entity-relation triples from Romanian legal documents
    Law-RoG is built entirely from LLM extraction; only 10 documents per domain were expert-validated (Section 3.3).
  • domain assumption BM25 with Romanian lemmatization retrieves the relevant legal subgraph
    GRAF samples the KG using BM25 top-10 terms and depth-1 BFS; if retrieval misses key facts, scores drop (Ablation Table 8).
  • domain assumption The exam labels are correct after OCR and manual inspection
    Dataset was collected from portals with OCR and manual checking; no expert legal validation is reported.
  • ad hoc to paper The number of correct answers for each question is known at test time
    Inference selects top |T_i| choices (Appendix D), an assumption that does not hold in real exams.
invented entities (1)
  • Law-RoG knowledge graph independent evidence
    purpose: Provides structured legal facts for GRAF's graph retrieval in Romanian.
    Built from CROL by Mixtral; only 10 documents per domain were checked by 5 human experts, so most of the 160k nodes and 320k edges lack direct validation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GRAF: Graph Retrieval Augmented by Facts for Romanian Legal Multi-Choice Question Answering." pith.science (2026). https://pith.science/paper/PXC6DHYO

@misc{pith2026241204119,
  author       = {Pith},
  title        = {Pith review of: GRAF: Graph Retrieval Augmented by Facts for Romanian Legal Multi-Choice Question Answering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PXC6DHYO}},
  note         = {Machine review of arXiv:2412.04119}
}
read the original abstract

Pre-trained Language Models (PLMs) have shown remarkable performances in recent years, setting a new paradigm for NLP research and industry. The legal domain has received some attention from the NLP community partly due to its textual nature. Some tasks from this domain are represented by question-answering (QA) tasks. This work explores the legal domain Multiple-Choice QA (MCQA) for a low-resource language. The contribution of this work is multi-fold. We first introduce JuRO, the first openly available Romanian legal MCQA dataset, comprising three different examinations and a number of 10,836 total questions. Along with this dataset, we introduce CROL, an organized corpus of laws that has a total of 93 distinct documents with their modifications from 763 time spans, that we leveraged in this work for Information Retrieval (IR) techniques. Moreover, we are the first to propose Law-RoG, a Knowledge Graph (KG) for the Romanian language, and this KG is derived from the aforementioned corpus. Lastly, we propose a novel approach for MCQA, Graph Retrieval Augmented by Facts (GRAF), which achieves competitive results with generally accepted SOTA methods and even exceeds them in most settings.

Figures

Figures reproduced from arXiv: 2412.04119 by the authors.

Figure 1
Figure 1. General topic graph for CROL. Red entities encompass the main legal branches, yellow entities rep￾resent topics, and grey entities are associated keywords. scenario, the answer can either be explicitly found in the provided context, in which case a model has to identify the location at which the answer oc￾curs (Zaib et al., 2021), or infer the correct answer based on understanding and interpretation of the context a… view at source ↗
Figure 2
Figure 2. GRAF procedure overview. and Jones, 1976) to select the top k entities in the knowledge graph. We proceed to select their vicinity with a breadth-first search for a limited depth. We also limit the maximum number of entities retrieved during this stage. In our work, we use a depth of 1, select the top 10 entities, nodes, and edges from the KG, and limit the selection to no more than 50 distinct entities. KG Encoding… view at source ↗
Figure 3
Figure 3. A showcase of how GRAF works on a sample question from the JuRO dataset translated to English. First, we construct the claim graph using an LLM that extracts the entities (nodes) and relations between them (edges). Based on the sub-KG extracted from Law-RoG and the claim graph, we determine that the entities “Self-defense” and “Unauthorized Access” have the best alignment with the entities in choice “C”, thus, it is… view at source ↗
Figures from the paper (9 more)
Figure 5
Figure 5. Figure 5: Comparative performances of different back [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: On the left, the number of samples from each law category in the JuRO dataset. On the right, the class [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 7
Figure 7. Figure 7: The distribution of top 30 topics in the JuRO dataset. [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: The distribution of top 20 topics in the CROL dataset. [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]
Figure 9
Figure 9. Figure 9: The token distribution on the JuRO dataset for Flan-T5 and Mistral tokenizers. [PITH_FULL_IMAGE:figures/full_fig_p020_9.png]
Figure 10
Figure 10. Figure 10: Per topic average pairwise percentage agreement scores when employing Llama-3.1 8B Instruct, FLAN [PITH_FULL_IMAGE:figures/full_fig_p033_10.png]
Figure 11
Figure 11. Figure 11: Accuracy computed for samples in the top 13 topics from the JuRO dataset, for every LLM. Higher is [PITH_FULL_IMAGE:figures/full_fig_p033_11.png]
Figure 12
Figure 12. Figure 12: Per-topic question difficulty on the JuRO dataset relative to LLM performance using the z-score [PITH_FULL_IMAGE:figures/full_fig_p034_12.png]
Figure 13
Figure 13. Figure 13: The dependency between accuracy, question difficulty (as z-score), model, and topic size. Larger [PITH_FULL_IMAGE:figures/full_fig_p035_13.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 11 canonical work pages

  1. [8]

    InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 22266–22275

    Interpretable long-form legal question answer- ing with retrieval-augmented large language models. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 22266–22275. Jarana Manotumruksa, Jeff Dalton, Edgar Meij, and Em- ine Yilmaz. 2020. Crossbert: a triplet neural archi- tecture for ranking entity properties. InProceedings of ...

    arXiv 2020

  2. [9]

    InProceedings of the ACM/IEEE Joint Confer- ence on Digital Libraries in 2020, JCDL ’20, page 385–388, New York, NY , USA

    Towards an open platform for legal informa- tion. InProceedings of the ACM/IEEE Joint Confer- ence on Digital Libraries in 2020, JCDL ’20, page 385–388, New York, NY , USA. Association for Com- puting Machinery. Vasile P˘ai¸ s, Radu Ion, Elena Irimia, Verginica Barbu Mititelu, Valentin Badea, and Dan Tufis, . 2024. Sys- tem for the anonymization of romani...

    work page 2020

  3. [10]

    Gerard Salton, Anita Wong, and Chung-Shu Yang

    Qa dataset explosion: A taxonomy of nlp resources for question answering and reading com- prehension.ACM Computing Surveys, 55(10):1–45. Gerard Salton, Anita Wong, and Chung-Shu Yang. 1975. A vector space model for automatic indexing.Com- munications of the ACM, 18(11):613–620. Priyanka Sen, Alham Fikri Aji, and Amir Saffari

    work page 1975

  4. [11]

    In Proceedings of the 29th International Conference on Computational Linguistics, pages 1604–1619

    Mintaka: A complex, natural, and multilin- gual dataset for end-to-end question answering. In Proceedings of the 29th International Conference on Computational Linguistics, pages 1604–1619. Atreya Shankar, Andreas Waldis, Christof Bless, Maria Andueza Rodriguez, and Luca Mazzola. 2023. Pri- vacyglue: A benchmark dataset for general language understanding ...

    work page 2023

  5. [12]

    R˘azvan-Alexandru Sm˘adu, Ion-Robert Dinic˘a, Andrei- Marius Avram, Dumitru-Clementin Cercel, Florin Pop, and Mihaela-Claudia Cercel

    Towards expert-level medical question an- swering with large language models.arXiv preprint arXiv:2305.09617. R˘azvan-Alexandru Sm˘adu, Ion-Robert Dinic˘a, Andrei- Marius Avram, Dumitru-Clementin Cercel, Florin Pop, and Mihaela-Claudia Cercel. 2022. Legal named entity recognition with multi-task domain adaptation. InProceedings of the Natural Legal Lan- g...

    arXiv 2022

  6. [14]

    InJSAI International Symposium on Artificial Intelligence, pages 49–65

    Improving vietnamese legal question– answering system based on automatic data enrich- ment. InJSAI International Symposium on Artificial Intelligence, pages 49–65. Springer. Jiajia Wang, Jimmy Xiangji Huang, Xinhui Tu, Junmei Wang, Angela Jennifer Huang, Md Tahmid Rahman Laskar, and Amran Bhuiyan. 2024. Utilizing bert for information retrieval: Survey, ap...

    arXiv 2024

  7. [17]

    This, together with the previous baseline, achieves the best results among the baselines

    adaptation method, which was experimen- tally shown to match the performance of classic full parameter fine-tuning. This, together with the previous baseline, achieves the best results among the baselines. We opt for the LoRA strategy, since our computational resources would not allow a full arameter fine-tuning of all our proposed LLMs. G Language Models...

    work page 2020

  8. [18]

    implemented in thebitsandbyteslibrary. Al- though this model is relatively large, we utilized it in the KG construction and claim extraction pro- cess as a trustworthy means, which is more likely to correctly extract entities and relations. More lightweight solutions can be built by training a smaller language model for this task or by distilling (Hinton,...

    work page arXiv 2015

Show all 18 references
  1. [1975]

    penal” com- pared to other words, indicating a possible preva- lence of penal-related content in the dataset. More- over, the terms such as “case

    for JuRO dataset, calculated using the fol- lowing formula: score(t, C) =f(t,C) |{w|d∈C,w∈d}|log |C| |{d|d∈C and t∈d}|(10) where: • the current term for which we compute the score is denoted byt; •C is the corpus of documents, each document containing multiple words; •f(t, C) ...

    2020

  2. [2014]

    InNew Frontiers in Artificial Intelligence, pages 199– 213, Cham

    Answering yes/no questions in legal bar exams. InNew Frontiers in Artificial Intelligence, pages 199– 213, Cham. Springer International Publishing. Mi-Young Kim, Ying Xu, Yao Lu, and Randy Goebel

  3. [2015]

    All BERT models were fine-tuned on all parameters

    optimizer via vanillaPyTorch. All BERT models were fine-tuned on all parameters. LLMs were fine-tuned for 50 epochs using theTrainer APIprovided by thetransformerslibrary using a learning rate of 10−7, AdamW optimizer, LoRA (Hu et al., 2022) alpha of 32, LoRA rank 64, and 2 wa...

    2020

  4. [2017]

    InNew Frontiers in Arti- ficial Intelligence, pages 299–313, Cham

    Question answering of bar exams by paraphras- ing and legal text analysis. InNew Frontiers in Arti- ficial Intelligence, pages 299–313, Cham. Springer International Publishing. Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In3rd Inter- nati...

    2015

  5. [2018]

    Thi-Hai-Yen Vuong, Ha-Thanh Nguyen, Quang-Huy Nguyen, Le-Minh Nguyen, and Xuan-Hieu Phan

    Graph Attention Networks.International Con- ference on Learning Representations (ICLR). Thi-Hai-Yen Vuong, Ha-Thanh Nguyen, Quang-Huy Nguyen, Le-Minh Nguyen, and Xuan-Hieu Phan

  6. [2020]

    InProceedings of the 28th International Conference on Computational Lin- guistics, pages 988–998, Barcelona, Spain (Online)

    Answering legal questions by learning neural attentive text representation. InProceedings of the 28th International Conference on Computational Lin- guistics, pages 988–998, Barcelona, Spain (Online). International Committee on Computational Linguis- tics. Mi-Young Kim, Randy ...

    2015

  7. [2021]

    InPro- ceedings of the International Conference on Recent Advances in Natural Language Processing (RANLP 2021), pages 92–101

    Pyeurovoc: A tool for multilingual legal docu- ment classification with eurovoc descriptors. InPro- ceedings of the International Conference on Recent Advances in Natural Language Processing (RANLP 2021), pages 92–101. Ngo Xuan Bach, Le Thi Ngoc Cham, Tran Ha Ngoc Thien, and T...

    2021 arXiv

  8. [2022]

    InFindings of the Association for Computational Linguistics: EMNLP 2022, pages 2518–2532

    Banglarqa: A benchmark dataset for under- resourced bangla language reading comprehension- based question answering with diverse question- answer types. InFindings of the Association for Computational Linguistics: EMNLP 2022, pages 2518–2532. Yunfan Gao, Yun Xiong, Xinyu Gao, ...

    2022 arXiv

  9. [2023]

    Knowledge distillation of large language mod- els.arXiv preprint arXiv:2306.08543. Neel Guha, Julian Nyarko, Daniel Ho, Christopher Ré, Adam Chilton, Aditya K, Alex Chohlas-Wood, Austin Peters, Brandon Waldon, Daniel Rockmore, Diego Zambrano, Dmitry Talisman, Enam Hoque, Faiz ...

    2023 arXiv

  10. [2024]

    InFindings of the Association for Computational Linguistics: EMNLP 2024, pages 4531–4541, Miami, Florida, USA

    RoQLlama: A lightweight Romanian adapted language model. InFindings of the Association for Computational Linguistics: EMNLP 2024, pages 4531–4541, Miami, Florida, USA. Association for Computational Linguistics. Phong-Khac Do, Huy-Tien Nguyen, Chien-Xuan Tran, Minh-Tien Nguyen,...

    2024 arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.