BaZi2500

BaZi2500 is a Chinese four-option multiple-choice benchmark for evaluating how well language models handle traditional Chinese Bazi (八字, Four Pillars) knowledge and the structured application of its rules to worked cases.

It contains 2,492 items in two task families:

Task family	task value	Items	Categories	What it tests
Theory (命理知识题)	mingli_question	1,454	14	Recall and application of codified rules (Five Elements, Ten Gods, combinations and clashes, useful/adverse gods, void, nayin, …)
Case (命例题)	mingli_example	1,038	11	Reading a given chart (four pillars + gender, sometimes annual/luck cycles) and inferring a described life outcome or situation (career, wealth, health, romance, family, education, personality, timing, chart construction, luck pillars, misfortune)

Random baseline accuracy is 25%. The primary metric is exact-match accuracy on the single gold letter.

Scope of what this measures. The benchmark measures agreement with a released answer key about a traditional Chinese metaphysical system. It makes no claim about the real-world predictive validity of Bazi practice, and it must not be used to produce or justify advice about real people's health, finances, relationships, or life decisions.

中文摘要（简要）：BaZi2500 是一个中文四选一 benchmark，用于评测大模型对中国传统八字命理 知识规则与命例推演的掌握程度，共 2492 题（理论题 1454 / 命例题 1038，25 个子类）。 随机基线 25%，主指标为答案字母精确匹配。数据为原始 3000 题的只删不改精炼版。 本数据集只衡量「与所发布答案键的一致程度」，不代表八字命理对现实世界具有预测效度， 不得用于替真人做健康、财务、感情或人生决策。详见下文中英对照字段表与局限说明。

---

Release status

BaZi2500 is publicly available on Hugging Face and is the recommended benchmark release for this paper.

Item	Status
Dataset availability	Public: MonsterPPPPP/BaZi2500
Dataset license	CC BY-NC 4.0
Source redistribution rights	Confirmed by the authors; the classical and traditional source materials have no modern copyright restrictions
Privacy	Human review completed; privacy issues have been addressed and confirmed by the authors
Expert human adjudication of gold answers	Not comprehensively completed; the key is model-generated and model-verified
Contact	lijiulin@18trees.com

For source information and correction or takedown requests, see RIGHTS_AND_PROVENANCE.md.

Dataset Details

• Curated by: Jiulin Li and Ping Huang
• Shared by: MonsterPPPPP
• Language (NLP): Chinese (zh, Simplified)
• License: CC BY-NC 4.0 — Attribution-NonCommercial 4.0 International
• Repository (code, analysis, paper): https://github.com/MonsterPPPP/bazi-qa-benchmark
• Paper: Knowing the Rules, Applying the Rules: Evaluating Language Models on Traditional Chinese Bazi (arXiv:2610.05682)
• Paper PDF: https://arxiv.org/pdf/2610.05682
• Project Page: https://monsterpppp.github.io/bazi-qa-benchmark/
• Contact: lijiulin@18trees.com
• Funding: This work was supported by Beijing Liuyi Guanhua Technology Co., Ltd.
• Competing interests: The authors declare no competing interests.

Dataset Sources

Both task families were constructed from a pipeline run over digitised Chinese Bazi source materials (see Dataset Creation). The authors confirm that the classical and traditional source materials have no modern copyright restrictions and that the dataset may be publicly redistributed in its current form under CC BY-NC 4.0.

Uses

Direct Use

Zero-shot or few-shot evaluation of language models on Chinese multiple-choice questions about Bazi theory and case analysis.

from datasets import load_dataset

ds = load_dataset("MonsterPPPPP/BaZi2500", split="test")
print(ds)            # Dataset({features: [...], num_rows: 2492})
print(ds[0])

theory = ds.filter(lambda r: r["task"] == "mingli_question")   # 1454
case   = ds.filter(lambda r: r["task"] == "mingli_example")    # 1038

The release is a single test split: all 2,492 items are evaluation items. There is no training split, and no item appears in a training set distributed here.

Reference prompt. The items were evaluated with the prompt reproduced verbatim in evaluation/PROMPT.md; a local scorer for {id, prediction} prediction files is in evaluation/evaluate_predictions.py.

Out-of-Scope Use

• Fortune-telling, divination, or life advice. Nothing here supports claims about a real person's future. Do not deploy models trained or prompted with this data to advise people on health, money, marriage, or major decisions.
• Claims of real-world predictive validity for Bazi, or of cultural authority. The benchmark scores agreement with an answer key, nothing more.
• Definitive model rankings. On the reference six-model panel the top three are statistically indistinguishable, and the item set is model-informed (see Known issues). "Model X is best at Bazi" is not supported by this data.
• Cultural or demographic generalisation. Category balance is an editorial choice; it does not imply balanced coverage of sources, schools, or regions.
• Training data for advice-giving systems about people's real lives.

Dataset Structure

data/test.jsonl — one JSON object per line, 2,492 lines, UTF-8, LF.

Field	Type	Description
id	string	Stable item id. Prefix mq4_ = Theory, me4_ = Case. Unchanged from the upstream release, so items join to upstream analysis artifacts by id.
task	string	mingli_question (Theory) or mingli_example (Case).
question	string	Question stem. Case items begin with a 【命例信息】 block (gender and the four pillars, sometimes annual/luck cycles), then 【问题】 and the question itself.
choices	list[string]	Exactly four options in A–D order: choices[0] = A … choices[3] = D.
answer	string	Gold answer letter: "A", "B", "C", or "D".
answer_index	int64	0–3, position of the gold answer in choices (so choices[answer_index] is the gold option text). JSON integers load as int64.
category	string	Chinese category name — 14 for Theory, 11 for Case.
structure_tag	string	Theory: which_choice, direct_query, reasoning, list_selection, judgement. Case: always case_context.
fact_or_theory	string	Case only (理论 = theory-grounded, 事实 = fact/observation). Empty string for Theory items.
gender	string	Case only (male / female). Empty string for Theory items.

Empty-string convention. Case-only fields hold "" on Theory rows rather than null, so the column types are unambiguous. Filter with ds.filter(lambda r: r["fact_or_theory"] != "").

Example row:

{"id": "me4_000001", "task": "mingli_example", "question": "【命例信息】\n性别：男\n八字：年柱丁巳；月柱壬寅；日柱辛酉；时柱乙未\n【问题】根据原命局，此命的事业身份最可能是什么？", "choices": ["生意人", "做官的", "艺术家", "普通的平民百姓"], "answer": "D", "answer_index": 3, "category": "事业", "structure_tag": "case_context", "fact_or_theory": "理论", "gender": "male"}

Categories

Theory (14 categories, task = mingli_question):

ID	English	中文	n
T01	Five Elements	五行	104
T02	Stems & Branches	天干地支	104
T03	Ten Gods	十神	104
T04	Ten Gods Relations	十神关系	104
T05	Combination & Clash	合冲刑害破	104
T06	Strength & Weakness	旺衰强弱	104
T07	Pattern Basics	格局基础	104
T08	Useful & Adverse Gods	用神忌神	104
T09	Shensha Basics	神煞基础	104
T10	Nayin	纳音	104
T11	Kongwang (Void)	空亡	104
T12	Twelve Stages	十二长生	104
T13	Annual & Luck Cycles	流年大运基础	102
T14	Six Relations	六亲	104

Case (11 categories, task = mingli_example):

ID	English	中文	n
C01	Career	事业	96
C02	Education	学业	96
C03	Wealth	财富	96
C04	Romance	感情	96
C05	Health	健康	96
C06	Personality	性格	96
C07	Year Timing	流年应期	96
C08	Chart Construction	排盘原局	96
C09	Luck Pillars	排大运	78
C10	Misfortune & Legal	灾厄官非	96
C11	Family Relations	六亲	96

T13 holds 102 items because two items in that category were confirmed duplicate/conflicting and were removed; C09 holds all 78 items that exist in that category in the source release.

Answer distribution

Answer	A	B	C	D
Items	646	607	624	615

Roughly balanced; random baseline is 25%.

Field mapping from the upstream release

release/clean-release/*.jsonl (upstream) → data/test.jsonl (this repo):

Upstream	Here
input.question	question
input.options = {"A": …, "D": …}	choices = […, …] (A–D order)
target.answer	answer (plus new answer_index)
meta.category / meta.structure_tag / meta.fact_or_theory / meta.gender	same names, flattened to top level
split: "publish"	dropped; the real split is test
meta.subcategory	dropped — see below

No question text, option text, or gold answer letter was modified. The transformation is purely structural. audit/manifest.json records "text_modified": false and the SHA-256 of every file.

meta.subcategory was dropped. In the upstream data this field is present on Theory items only, and its value equals the gold answer letter in 1,454 of 1,454 Theory items (distribution A 364 / B 363 / C 365 / D 362, identical to the answer distribution). The name says "subcategory" but the content is a copy of the gold label, so publishing it under that name would create a misleading, fully redundant column. To reconstruct it: subcategory = answer.

Dataset Creation

Curation Rationale

Bazi is a rule-governed traditional Chinese system with a large, stable vocabulary (Ten Gods, Five Elements, combinations/clashes, useful gods, …) that is well suited to multiple-choice evaluation, yet it is almost absent from existing Chinese benchmarks. BaZi2500 targets two distinct capabilities:

• Theory — retrieval and application of codified rules, where models are expected to do well.
• Case — applying those rules to a specific chart to infer a described outcome, which requires multi-step reasoning over a structured context.

Keeping the two families separate is deliberate: they behave very differently, and a pooled score hides that.

Source Data

Data collection and processing. Items were produced by a staged pipeline over digitised Chinese Bazi books:

stage0 raw books  →  stage1 chapter texts  →  stage2 fine text  →  stage3 QA data  →  stage4 release

• Admission required direct textual evidence, extractable professional knowledge, a scorable target, and a category boundary consistent with a frozen rulebook; generic discussion, book introductions, keyword-only passages, and case labels without conditions/conclusions were rejected.
• A three-layer label gate (task gate → category review → per-label verification) assigned each knowledge unit to a category.
• Generation, verification, and near-duplicate detection used deepseek-chat (temperatures 0.1 / 0.2 / 0.3 / 0.1; up to three candidates per knowledge unit; near-duplicate threshold 0.92). Items were structurally validated to have exactly four non-empty, distinct options.
• Difficulty and discrimination used in the refinement step were computed from the responses of six model endpoints on the 3,000-item original release.

Who produced the source data? The underlying books are the work of their named authors and publishers, whose identities are recorded upstream in a book registry. The authors have confirmed redistribution rights for this release. Item stems, options, and gold answers are machine-generated text derived from those books and then machine-verified; they are not verbatim quotations from a single source.

From 3,000 to 2,492 items (deletion-only refinement)

The released set is a deletion-only derivative of an archived 3,000-item release (1,500 Theory + 1,500 Case). 508 ids were removed:

Removal band	Items
All six reference models correct (all_correct)	477
Five of six correct (easy_5)	30
All six wrong (all_wrong)	1
Total	508

Four removals were forced data-defect cases (two items with identical stems but conflicting gold answers; two of three near-duplicate items) and overlap the bands above.

The priority order was chosen using the paired McNemar statistic z ≈ |b − c| / √(b + c), which depends only on discordant pairs. Items that every model answers correctly (or every model answers wrongly) are concordant pairs: removing them changes no b or c and therefore cannot change any pairwise comparison. The refinement deleted that band first, then capped each (Theory, category) at 104 items and each (Case, category) at 96 items, sampling within difficulty bands with a fixed seed (20260922).

The complete rules are in audit/CLEANING_SPEC.md; per-category before/after counts are in audit/cleaning_stats.csv; every removed id with its reason and its difficulty/discrimination metrics is in audit/removed_items.csv.

Annotations

Gold answers are model-generated and model-verified. There is no human expert annotation set and no inter-annotator agreement report for this release: the two conflicting items were withdrawn pending human adjudication, and that adjudication has not been performed. The queue of candidate items for expert review (all-wrong items, cue-heavy items, reasoning-reversal items) is a queue, not a record of decisions.

Personal and Sensitive Information

The authors have manually reviewed and addressed privacy in the public release and confirm that privacy handling is complete. Case items contain gender, four-pillar charts, and personal-style narratives drawn from traditional source materials. Do not attempt to re-identify individuals. For correction or removal requests, contact lijiulin@18trees.com; see RIGHTS_AND_PROVENANCE.md.

Bias, Risks, and Limitations

Known issues in the release, measured on the 2,492 items:

1. The item set is model-informed, so it is not a clean held-out benchmark. Difficulty bands, discrimination, the removal priority, and the retention choice among near-duplicates all use the six reference models' responses. Selection can flatter or penalise those specific models. Interval estimates and significance tests computed on this set are post-selection descriptions, not confirmatory tests. On this set, 11 of 15 Holm-corrected pairwise McNemar comparisons are detected as different; the non-rejections include every comparison among the top three models, which is not an equivalence claim.
2. Length cue. The gold option is uniquely the longest in 564 of 2,492 items (22.63%) — Theory 278/1,454 (19.12%), Case 286/1,038 (27.55%). This was deliberately left unfixed (see audit/CLEANING_SPEC.md §6) so the released items match the ones evaluated, but a model can exploit it.
3. Repeated charts. The 1,038 Case items reuse only 249 distinct charts; the most repeated chart appears 25 times with different questions. Items are not independent, and naive per-item significance tests overstate effective sample size.
4. Unverified gold. Gold answers lack comprehensive expert adjudication, so some "model errors" may be key errors. Items every reference model got wrong are not necessarily valid hard items.
5. Source dependence and contamination. Items descend from a limited set of digitised books, so a model pretrained on those books may have seen the underlying material. Because the full item text and answer key are public, this dataset is trainable and future results on it are not contamination-free.
6. Cultural and scope limits. The content reflects the schools and editorial choices of the source books, not "Chinese metaphysics" as a whole. Nothing here validates Bazi as a predictive practice.

Recommendations

• Report Theory and Case separately; the pooled number hides the main effect.
• Never report a single "best model" from this set without the caveat in item 1; prefer paired tests on an independently held-out item set.
• Treat the released key as authoritative for scoring but not as ground truth about the world.
• If you use the data for anything user-facing, add your own safety layer: this content is about fate, health, and death, and is easy to present as advice.
• Cite the specific version (see audit/manifest.json hashes) because the upstream release can be regenerated and re-refined.

Citation

Paper

@misc{li2026bazi2500,
  title = {Knowing the Rules, Applying the Rules: Evaluating Language Models on Traditional Chinese Bazi},
  author = {Li, Jiulin and Huang, Ping},
  year = {2026},
  eprint = {2610.05682},
  archivePrefix = {arXiv},
  primaryClass = {cs.CL},
  url = {https://arxiv.org/abs/2610.05682}
}

Dataset

Cite this public dataset as follows.

BibTeX:

@misc{bazi2500_2026,
  title        = {BaZi2500: A Chinese Multiple-Choice Benchmark for Traditional
                  BaZi Knowledge and Case Reasoning},
  author       = {Li, Jiulin and Huang, Ping},
  year         = {2026},
  howpublished = {Hugging Face dataset},
  url          = {https://huggingface.co/datasets/MonsterPPPPP/BaZi2500},
  note         = {2,492 items (1,454 theory, 1,038 case); deletion-only
                  refinement of a 3,000-item release}
}

APA:

Li, J., & Huang, P. (2026). BaZi2500: A Chinese multiple-choice benchmark for traditional BaZi knowledge and case reasoning [Data set]. Hugging Face. https://huggingface.co/datasets/MonsterPPPPP/BaZi2500

Glossary

Term	Meaning
八字 (BaZi)	"Eight characters": the year, month, day, and hour pillars of a birth moment, each a stem–branch pair.
四柱 (sizhu)	The four pillars (year/month/day/hour) that make up a chart.
天干 / 地支	The ten Heavenly Stems and twelve Earthly Branches that combine into 60 pillar pairs.
日主 / 日干	The day stem, treated as "the self" in a chart.
十神 (Ten Gods)	Ten relational roles a stem/branch plays relative to the day master (e.g. 正官/七杀 officer, 正财/偏财 wealth, 正印/偏印 resource, 食神/伤官 output, 比肩/劫财 peer).
五行	The Five Elements (wood, fire, earth, metal, water) and their generating/controlling relations.
合冲刑害破	Combination, clash, punishment, harm, and destruction relations among branches.
用神 / 忌神	The element(s) beneficial ("useful god") or harmful ("adverse god") to a chart.
大运 / 流年	Ten-year luck pillars and annual cycles layered onto a chart.
神煞	Named auxiliary "symbolic stars" marking specific chart configurations.
空亡	"Void": branches absent from the decade cycle a pillar belongs to.
纳音	The "sounding" element name assigned to each 60-cycle pillar pair.
理论 / 事实	In fact_or_theory: whether the item's premise is stated as doctrine (理论) or as an observed fact (事实).

More Information

• Repository, evaluation code, and paper: https://github.com/MonsterPPPP/bazi-qa-benchmark
• Project Page: https://monsterpppp.github.io/bazi-qa-benchmark/
• Cleaning specification (full rules and reproduction): audit/CLEANING_SPEC.md
• Removal manifest (all 508 removed ids with reasons): audit/removed_items.csv
• File hashes and provenance record: audit/manifest.json
• Evaluation prompt and scorer: evaluation/PROMPT.md, evaluation/evaluate_predictions.py
• Rights, provenance, and takedown: RIGHTS_AND_PROVENANCE.md

Dataset Card Authors

Jiulin Li and Ping Huang — card compiled from the upstream release artifacts and audit records.

Dataset Card Contact

lijiulin@18trees.com