VISAI-AI/nitibench-plus

NitiBench+ public release, version 1.0 (2026-11-01)

NITIBENCH-PLUS CANARY 7f3c1b2e-4a9d-4e61-9c58-2d0b6f8a1e47

NitiBench+ is a benchmark of Thai legal question answering built from decided cases of the Supreme Court, the Supreme Administrative Court and the Constitutional Court. Every question comes with the court's verdict, the court's reasoning and the statute sections the court relied on, each section identified by the version in force on the decision date and paired with the current version. The live leaderboard is at benchmark.fourcorners.law. The technical report NitiBench+: A Temporal Thai Legal Question Answering Benchmark Across 17 Domains (arXiv:2611.03576) describes the benchmark, the two splits and the leaderboard.

What this release contains

• validation_split.jsonl: 498 cases, 1,228 sub-questions. One JSON object per line. Every record carries the canary string in canary. For each case: id, court, decision number (source_id), decision date, domain, drift band, title and the facts of the case (fact_pattern, Thai). For each sub-question: the question, its kind (question_type: ruling for a question the court itself answered, or governing_provisions, legal_consequences, provision_interaction, lower_court_departure for the four added kinds described in Section 4.1 of the report), the court's verdict and reasoning (the gold answer), and the gold citations with law code, section label, the version in force on the decision date (version_uri, valid_from, valid_to) and the current version (current_uri). 130 of the 3,665 gold citations point to instruments the statute corpus does not hold (notifications, orders, some decrees); they carry the instrument's name in law and no identifier, and they are left out of Citation F1.
• gold_sections.jsonl: the text of every gold section in both versions, one record per section version with the canary string in canary, 3,253 section versions from the dated statute corpus (corpus date 2026-09-01), every identifier of the split resolved. This is what the Gold-cited setting feeds the model.
• grading/: the Thai Bar Council rubric, the judge prompt and the citation-matching rules, written out so that a score computed elsewhere means the same thing as ours.

The case texts come from published court decisions. In this release the names of private individuals and of officials named in person are replaced by their initials in the style of Thai law reports (นาย ส., นางสาว จ.), the same initial for the same person within a case, and citizen ID numbers, personal phone numbers, bank account numbers, vehicle plates and house numbers are masked. Nationally known public figures keep their names. This touches 210 of the 498 cases and changes no question, verdict, citation or score. Requests about a specific case go to the contact below.

• composition.json: how the split was drawn and what it holds.

Loading

import json

cases = [json.loads(l) for l in open("validation_split.jsonl", encoding="utf-8")]
sections = {r["version_uri"]: r for r in map(json.loads, open("gold_sections.jsonl", encoding="utf-8"))}

What is held out

The live leaderboard at benchmark.fourcorners.law runs on a test split of 500 cases drawn by the same recipe. It ranks every model in three settings with the cost per item and the hallucinated citations, and it takes new models as they are released. The test split is never published, neither its ids nor its text. A further 672 cases are kept in reserve to replace the test split if it leaks. The statute pool (105,402 sections, 394,943 dated versions), its index, the retrieval service and the evaluation harness are not released.

A model joins the leaderboard when we run it on the test split. Model makers who want a model on the board write to the FourCorners Team.

How the splits were drawn

The benchmark has 1,670 cases in 17 domains. The test split takes 29 or 30 cases per domain (30 in the seven largest domains) and, within a domain, the drift bands in the domain's own proportion, one case per shared-question cluster, in a deterministic order (SHA-256 of a salt and the case id). The validation split is drawn by the same rule from the cases outside the test split, with the same per-domain quota; Securities & Finance has 27 cases instead of 29 because only 27 remained. No case in this release shares a question cluster with a test case.

Domain	Cases	No drift	Amended	Repealed
Civil & Commercial	30	13	13	4
Criminal	30	3	22	5
Family & Succession	30	12	18	0
Intellectual Property	30	10	13	7
Labor	30	11	16	3
Land & Property	30	11	11	8
Tax & Revenue	30	5	12	13
Administrative Acts	29	13	12	4
Administrative Contracts	29	11	12	6
Narcotics	29	1	1	27
Civil Procedure	29	5	20	4
Public Law	29	8	8	13
Public Personnel	29	11	8	10
Natural Resources & Licensing	29	5	13	11
State Liability	29	13	8	8
Traffic & Transport	29	4	22	3
Securities & Finance	27	7	14	6

Drift bands over the split: 143 no drift, 223 amended, 132 repealed. Courts: 368 Supreme Court, 116 Supreme Administrative Court, 14 Constitutional Court.

The two splits agree

For the 13 open-weight models that ran all 1,670 cases at three samples, the score on the test split and the score on the validation split agree in every setting. The table gives the correlation over the 13 models and the difference between a model's two scores on the 10-point scale.

Setting	Pearson	Spearman	Mean absolute difference	Largest difference
Parametric	0.999	0.967	0.05	0.12
Gold-cited In-force	1.000	0.995	0.05	0.11
RAG In-force	0.999	0.989	0.05	0.16

A model whose validation score runs far ahead of its test score is the signature of training on the released data.

Settings a user can run from this release

• Parametric: the model answers from the facts and the question alone.
• Gold-cited In-force: the model also receives the text of the gold sections in the version in force (gold_sections.jsonl, join on version_uri).
• Retrieval settings need the statute pool and are run by the FourCorners Team only.

Canary

Every file of this release carries the string NITIBENCH-PLUS CANARY 7f3c1b2e-4a9d-4e61-9c58-2d0b6f8a1e47. Please filter documents containing it out of training corpora. The test split carries no canary because it is not published.

License and citation

The court decisions are Thai government documents and are in the public domain under Section 7 of the Copyright Act B.E. 2537. The annotations of this release (fact patterns, questions, gold answers, versioned citations) are licensed under CC BY 4.0. If you use this release, cite the technical report NitiBench+: A Temporal Thai Legal Question Answering Benchmark Across 17 Domains (arXiv:2611.03576).

@misc{akarajaradwong2026nitibenchtemporalthailegal,
      title={NitiBench+: A Temporal Thai Legal Question Answering Benchmark Across 17 Domains}, 
      author={Pawitsapak Akarajaradwong and Sarana Nutanong and Chompakorn Chaksangchaichot},
      year={2026},
      eprint={2611.03576},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2611.03576}, 
}

Contact: Pawitsapak Akarajaradwong, pawitsapaka_visai@vistec.ac.th (FourCorners Team, VISAI AI).