| --- |
| license: apache-2.0 |
| language: |
| - zh |
| tags: |
| - benchmark |
| - evaluation |
| - taiwan |
| - zh-tw |
| - multiple-choice |
| - twinkle-eval |
| task_categories: |
| - question-answering |
| size_categories: |
| - 10K<n<100K |
| pretty_name: OpenTWBench exam sample (leaderboard set) |
| configs: |
| - config_name: default |
| data_files: |
| - split: test |
| path: data/tw-exam-bench-sample.parquet |
| --- |
| |
| # tw-exam-bench-sample |
|
|
| **The 22,021 questions the [opentwbench.ai](https://opentwbench.ai) leaderboard |
| actually scores.** A fixed, deterministic stratified sample of |
| [`OpenTWBench/tw-exam-bench`](https://huggingface.co/datasets/OpenTWBench/tw-exam-bench), |
| so that a score you compute locally is comparable to a score on the board. |
|
|
| ```python |
| from datasets import load_dataset |
| |
| ds = load_dataset("OpenTWBench/tw-exam-bench-sample", split="test") |
| ``` |
|
|
| ## Why this exists separately |
|
|
| The full suite is 255,854 questions. Running it for one model is a day on eight |
| B200s; running it for the 107 models on the board is not a thing anyone will do. |
| So the board scores a sample — and a sample is only useful if everyone is scored |
| on the *same* one. |
|
|
| This dataset is that sample, published so it can be checked rather than trusted. |
| Every model on the board answered these rows and no others. |
|
|
| ## How it was drawn |
|
|
| Per subject, the `min(150, n)` questions with the **smallest `sha256(question)`**. |
| No seed, no shuffle, no reference to difficulty — re-running the draw on the same |
| corpus reproduces it exactly. |
|
|
| ``` |
| 148 subjects x 150 22,200 rows |
| less rows on the withdrawal list -179 |
| ---------- |
| 22,021 rows |
| ``` |
|
|
| The withdrawal list is the 2,180 items whose stems depend on a figure the published |
| text does not carry. 61 subjects end up under 150 because of it; the smallest, |
| `electronics`, has 123. |
|
|
| Because selection is by content hash, it is independent of whether any model answers |
| correctly. A sample drawn this way can differ from the population only through |
| **composition**, and that difference has been measured rather than assumed. |
|
|
| ## What a score on this set means |
|
|
| | | Qwen3.8-27B | |
| |---|---:| |
| | Full corpus, 253,674 items | **82.19%** | |
| | This sample, same per-subject accuracies re-weighted by its composition | **80.1%** | |
|
|
| Equal weighting per subject moves weight off the large, high-scoring medical corpus |
| and onto the small legal and administrative subjects, where models are weakest. So |
| a score here sits about **2 points below** what a full-corpus run would report. |
|
|
| That offset is constant and applies to every model, so rankings are unaffected. |
| **Scores on this set are comparable to each other. They are not comparable to a |
| full-corpus number.** |
|
|
| ## Fields |
|
|
| | Field | Description | |
| |---|---| |
| | `row_id` | 0–22,020, this dataset's canonical row order | |
| | `qh` | `sha256(question)[:16]` — the key the board's result files use. **Not unique**, see below | |
| | `subject` | Subject slug, matching the config names in `tw-exam-bench` | |
| | `subject_zh` | 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_ce` | Exam year, 2012–2026 | |
| | `paper_id` / `q_no` | Originating paper and question number — together a unique key | |
| | `question`, `A`, `B`, `C`, `D` | Stem and option texts | |
| | `answer` | Gold option letter | |
|
|
| The row shape is the **Twinkle Eval MCQ** contract; the provenance columns ride |
| alongside and are ignored by the harness. |
|
|
| ## A trap worth knowing about |
|
|
| Taiwan's examinations reuse a stem verbatim across years with a fresh set of options. |
| `下圖為何種通氣模式？` is six different questions in mechanical ventilation between |
| 2014 and 2019, with three different answer keys. |
|
|
| In this sample, **119 stems stand for 247 separate items**. So: |
|
|
| - `qh` collides. Keying your results on it discards 128 real questions. |
| - Score **row-wise**, over all 22,021 rows. Use `row_id`, or `(paper_id, q_no)`. |
|
|
| ## Protocol |
|
|
| Scored with [Twinkle Eval](https://github.com/ai-twinkle/Eval): temperature 0, |
| `shuffle_options: true`, answer extracted from `\boxed{}`. Items a |
| model never answers count as wrong rather than being dropped — the denominator is |
| the dataset, not the subset that parsed. |
|
|
| ## Relationship to the other datasets |
|
|
| | Dataset | Rows | Use | |
| |---|---:|---| |
| | [`tw-exam-bench`](https://huggingface.co/datasets/OpenTWBench/tw-exam-bench) | 255,854 | the full suite, per-subject configs | |
| | **`tw-exam-bench-sample`** | **22,021** | what the leaderboard scores | |
| | [`tw-formosa-bench`](https://huggingface.co/datasets/OpenTWBench/tw-formosa-bench) | 1,696 | everyday and contemporary Taiwan | |
|
|
| ## Licence and source |
|
|
| Apache-2.0. Questions are open data published by Taiwan's **Ministry of Examination |
| (考選部)**. |
|
|
