Build from 考選部 open data

README.md

@@ -0,0 +1,218 @@
+---
+license: apache-2.0
+language:
+- zh
+tags:
+- benchmark
+- evaluation
+- taiwan
+- zh-tw
+- multiple-choice
+- twinkle-eval
+- refraction
+task_categories:
+- question-answering
+size_categories:
+- n<1K
+configs:
+- config_name: default
+  data_files:
+  - split: test
+    path: benchmark.csv
+---
+
+# Refraction (驗光學)
+
+A multiple-choice benchmark for **Refraction** — the subject 「驗光學」 as
+examined in Taiwan's national examinations, in Traditional Chinese. Built from the
+papers published as open data by the **Ministry of Examination (考選部)**.
+
+One dataset per **academic subject**: papers are grouped by the subject they examine,
+not by the professional category sitting the exam.
+
+Part of [**OpenTWBench**](https://huggingface.co/OpenTWBench) — an open evaluation suite for
+Taiwan-domain knowledge.
+
+## Overview
+
+| Property | Value |
+|---|---:|
+| Questions | **732** (deduplicated) |
+| Years | 2017–2026 |
+| Source papers parsed | 15 of 15 |
+| Format | 4-choice single-answer (A/B/C/D) |
+| Language | Traditional Chinese (zh-TW) |
+| Framework | Twinkle Eval compatible |
+| Field | optometry |
+
+## Fields
+
+The row shape is the **Twinkle Eval MCQ** contract — `question`, `A`–`D`, `answer` —
+with provenance columns alongside, which the harness ignores.
+
+| Field | Description |
+|---|---|
+| `question` | Question stem |
+| `A` / `B` / `C` / `D` | Option texts |
+| `answer` | Gold option letter |
+| `subject` | Full examination subject name (科目全名) |
+| `subject_group` | Examination category (類科組別) |
+| `exam_level` | Examination grade (等級分類) |
+| `exam_name` | Official title of the examination |
+| `year_roc` / `year_ce` | Exam year (ROC / CE calendar) |
+| `q_no` | Question number on the original paper |
+| `paper_id` | Originating paper identifier |
+| `n_source_papers` | How many papers published this identical item |
+
+## Examination Categories
+
+| 類科組別 | Questions |
+|---|---:|
+| 驗光生 | 732 |
+
+## Examination Levels
+
+| 等級分類 | Questions |
+|---|---:|
+| 專技普考 | 493 |
+| 相當普考 | 239 |
+
+## Usage
+
+### Twinkle Eval
+
+```bash
+twinkle-eval --init
+# point `dataset_paths` at a directory holding benchmark.jsonl, then:
+twinkle-eval --config config.yaml --export json csv
+```
+
+### datasets
+
+```python
+from datasets import load_dataset
+
+ds = load_dataset("OpenTWBench/tw-refraction-bench", split="test")
+print(ds[0]["question"], ds[0]["answer"])
+```
+
+## Construction
+
+1. **Index** — the Ministry of Examination open-data index
+   (`wHandExamQandA_CSV.ashx`), filtered to multiple-choice papers that publish an
+   answer key.
+2. **Retrieval** — question and answer-key PDFs fetched directly, rate-limited and
+   cached.
+3. **Parsing** — PDF text layer extracted with PyMuPDF; see the hazards below.
+4. **Filtering and deduplication** — see below.
+
+### Parsing hazards handled
+
+These silently corrupt naive extraction, so they are documented here:
+
+- Option markers live in the **EUDC Private Use Area** (`U+E18C`–`U+E190` = A–E)
+  rather than as literal letters; some older layouts use `A.` or `(A)` instead.
+- Text mixes **CJK Compatibility Ideographs**. Normalised with **NFC** — *not*
+  NFKC, which rewrites 「，」 to an ASCII comma.
+- Answer-key PDFs are **grids whose text layer comes out shuffled**: question
+  labels and answer letters arrive in separate runs, sometimes out of order.
+  Pairing by reading order misaligns the entire paper, so cells are paired **by
+  page coordinates** instead, with the column tolerance derived from the grid pitch.
+- The first answer cell of a row is often **glued to the row label** in the text
+  layer (`答案Ｃ`). Left unhandled, that cell goes missing and every label in the
+  row pairs with its right-hand neighbour — shifting the whole paper by one.
+- Question numbers must run **1..N monotonically**: a bare number inside a stem
+  (frequent in quantitative papers) would otherwise start a phantom question.
+- Answer letters may be **full-width** (`Ａ`) and `＃` marks a voided question.
+- Papers mix **single-select and multi-select** sections; multi-select keys are
+  multi-letter tokens (`ABDE`).
+
+Every paper is cross-checked against the question count its own answer sheet
+declares (`單選題數`/`複選題數`/`題數`); a mismatch rejects the paper rather than
+importing a silently shifted key.
+
+### Excluded
+
+| Excluded | Count |
+|---|---:|
+| Multi-select items | 0 |
+| Voided items (一律給分) | 8 |
+| English-language items | 0 |
+| Items referring to a figure or table | 0 |
+| Follow-up items in a 題組 (「承上題」) | 0 |
+| Items whose last option swallowed following text | 0 |
+| Items not in 4-choice form | 10 |
+| Items whose stem was displaced by table debris | 0 |
+| Items whose duplicate keys disagreed | 0 |
+| Items already in tw-legal-benchmark-v2 | 0 |
+| Papers rejected by the answer-count check | 0 |
+| Papers rejected as incompletely parsed | 0 |
+
+Questions already published in
+[`lianghsun/tw-legal-benchmark-v2`](https://huggingface.co/datasets/lianghsun/tw-legal-benchmark-v2)
+are removed **item by item**, not by dropping whole domains: several 法規 subjects
+(營建法規, 海巡法規, 郵政法規, 稅務法規) are examined inside otherwise non-legal
+categories, so excluding the category would discard its non-legal questions too.
+
+English items are detected from the **stem**, not by question number: the boundary
+of the English section moves between years. Judging the options too would discard
+the many Chinese questions that are answered in English terminology — 「下列何者可
+以抑制 ribonucleotide reductase 的活性？」 with options like `dATP` is a Chinese
+question about Taiwan-examined knowledge, and a zh-TW model is expected to answer
+it.
+
+Items that point at a figure, table or reading passage printed on the paper are
+removed — they are unanswerable from the text alone, and keeping them would make
+every model look worse for the same reason. So are the follow-up questions of a
+題組, which open with 「承上題」 and refer to a scenario set up by the previous
+question: the setup question is self-contained and stays, the follow-ups cannot
+be answered by anyone once lifted out of the paper.
+
+The last option on a paper has no following marker for the parser to stop at, so
+table debris or the next question's text can run into it. Option texts are short
+— the 99.9th percentile is 95 characters — so an option far outside that *and*
+several times longer than its siblings is a parse failure, and the item is
+dropped rather than shipped broken.
+
+### Deduplication
+
+The same paper is republished under several examination categories, so the same
+item recurs verbatim. Items are fingerprinted on their stem plus their sorted
+option set; **0** duplicate groups were collapsed,
+of which **0** carried disagreeing answer keys and
+were dropped entirely rather than guessed. `n_source_papers` records the
+multiplicity — **0** of the kept items were published more than once.
+
+**A benchmark built from this source without deduplication is inflated**, and its
+accuracy is skewed toward whichever subset happens to be republished most.
+
+## Known Properties
+
+- **Answer-position distribution**: A=153, B=208, C=188, D=183.
+  Shuffle options when evaluating (`shuffle_options: true` is the Twinkle Eval default).
+- **Contamination risk**: these are public past papers, widely discussed online and
+  very likely present in pretraining corpora. Absolute scores should be treated with
+  caution; the benchmark is most useful for **relative** comparison between models.
+- **Answer extraction dominates low scores.** If a model scores near zero, check the
+  unparsed rate before believing the number — a model answering 「最終答案：C」 scores
+  very differently under `box` and `pattern` extraction. Match the extraction method
+  to the model.
+- Item difficulty is set by the examination committee, not by us.
+
+## Licensing
+
+Questions and answer keys are official publications of the Ministry of Examination,
+Republic of China (Taiwan), released as open data. The packaging, parsing,
+deduplication and metadata in this repository are provided under Apache-2.0. Users
+intending redistribution should confirm the current terms published by 考選部.
+
+## Citation
+
+```bibtex
+@misc{opentwbench_refraction,
+  title  = {Refraction Benchmark (Taiwan)},
+  author = {Huang, Liang-Hsun},
+  year   = {2026},
+  url    = {https://huggingface.co/datasets/OpenTWBench/tw-refraction-bench}
+}
+```