README.md

metadata

language:
  - zh
  - en
language_bcp47:
  - zh-Hant
  - zh-Hans
  - en
license: mit
pretty_name: TW-LegalBench
size_categories:
  - 10K<n<100K
task_categories:
  - multiple-choice
  - text-generation
  - question-answering
tags:
  - legal
  - taiwan
  - benchmark
  - icail-2026
configs:
  - config_name: close_exam_tc
    data_files:
      - split: test
        path: close_exam/tc/test.parquet
  - config_name: close_exam_sc
    data_files:
      - split: test
        path: close_exam/sc/test.parquet
  - config_name: close_exam_en
    data_files:
      - split: test
        path: close_exam/en/test.parquet
  - config_name: open_exam
    data_files:
      - split: test
        path: open_exam/test.parquet
  - config_name: verdict_prediction
    data_files:
      - split: train
        path: verdict_prediction/train.parquet
      - split: test
        path: verdict_prediction/test.parquet

\

TW-LegalBench

Taiwan legal reasoning benchmark for large language models. Accepted to ICAIL 2026.

Configs

Config	Split	Rows	Description
close_exam_tc	test	16643	Multiple-choice questions from Taiwan legal exams (2020–2024), Traditional Chinese
close_exam_sc	test	16643	Same questions, Simplified Chinese translation
close_exam_en	test	16643	Same questions, English translation
open_exam	test	117	Bar-exam essay questions with official scoring rubrics
verdict_prediction	train / test	13790 / 535	Predict a Taiwan criminal verdict main text (主文) from case facts

Load

from datasets import load_dataset

# MCQ (Traditional Chinese)
mcq = load_dataset("feiyuehchen/TW-LegalBench", "close_exam_tc", split="test")

# Essay
essay = load_dataset("feiyuehchen/TW-LegalBench", "open_exam", split="test")

# Verdict prediction
verdict = load_dataset("feiyuehchen/TW-LegalBench", "verdict_prediction")
# → DatasetDict({"train": ..., "test": ...})

Metrics reported in the paper

• Close Exam — accuracy over 4 options (A–D).
• Open Exam — LLM-as-Judge coverage of the official scoring points, graded on a 3-point scale (yes / partial / no).
• Verdict Prediction — ROUGE-1/2/L (jieba tokenization) + multilingual sentence-transformers similarity + legal-validity heuristics (defendant / crime / sentence / law-article presence, hallucinated-law detection).

See the paper for full methodology and reported numbers.

Fields

close_exam_*

Top-level: question_id, question, options (list, A–D), correct_answer (string A–D), and a nested metadata struct:

metadata:
  year, exam_id, exam_category, professional_level, profession, subject,
  full_name, primary_key, source,
  is_law_related, jurisdiction,
  related_internal_law, related_international_law,
  laws_tc, laws_en

Some metadata fields are sparse by design — e.g. related_international_law is only populated when a hand-annotator flagged a question as citing an international instrument (~125 rows).

open_exam

Top-level: question_id, year, subject, subject_code, question_number, question_text, question_text_with_points, points, reference_law_names, scoring_points (list of struct with point_id / description / max_score), plus a nested metadata struct with exam_name / exam_type / question_type and derived flags.

Dropped: sub_questions, sub_questions_with_points, reference_laws (all 100% null in the source).

verdict_prediction

All 28 non-null upstream columns preserved. Paper task uses fact → main_result.

Dropped: 10 all-NaN pipeline leftovers (law_used, history_trial, history_res, link, warn_lawyer, per_result, is_murder, puni_murder, main_crime, indictment) plus the pandas Unnamed: 0 row-number artifact.

License

MIT for the code and annotation effort. The underlying exam questions are released under the respective copyright terms of the Taiwan Ministry of Examination; the verdict texts are public judicial records from the Judicial Yuan. Cite the paper if you use this benchmark.