Datasets:
File size: 4,283 Bytes
77ca2f1 9c732f3 77ca2f1 9c732f3 d3014e7 9c732f3 4bb6074 9c732f3 d3014e7 4bb6074 d3014e7 9c732f3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 | ---
license: cc-by-nc-sa-4.0
language:
- zh
task_categories:
- question-answering
pretty_name: TCMQA
size_categories:
- 10K<n<100K
tags:
- medical
- traditional-chinese-medicine
- tcm
- multiple-choice
- benchmark
configs:
- config_name: qa
default: true
data_files:
- split: train
path: data/qa/train-*
- config_name: doctor_annotated
data_files:
- split: train
path: data/doctor_annotated/train-*
---
# TCMQA
*Built by [TechTCM](https://techtcm.com/).*
A Traditional Chinese Medicine question-answering benchmark: **33,872 Chinese-language
multiple-choice questions** drawn from TCM licensing-exam material, plus a
**5,250-question subset answered by 102 licensed TCM practitioners**, giving a human
reference point for the same items a model is scored on.
## Configs
### `qa` (default) — 33,872 rows
| Field | Type | Notes |
|---|---|---|
| `id` | int64 | Stable question id. Not contiguous — ids are preserved from the source collection across cleaning. |
| `question` | string | Question stem, Chinese. |
| `choices` | struct `{A,B,C,D,E}` | Option text. **`E` is an empty string on 4-option questions** (12,193 rows); use `num_choices` rather than assuming five. |
| `num_choices` | int64 | 4 or 5. |
| `answer` | string | Correct option letters, sorted, e.g. `"B"` or `"BD"`. |
| `num_answers` | int64 | 1–5. |
| `is_multi_answer` | bool | True for 2,308 rows (6.8%). |
| `explanation` | string / null | Reference rationale where the source provided one (20,692 rows). |
| `topic` | string / null | Free-text topic label (23,779 rows, 5,738 distinct). Uncontrolled vocabulary — see Limitations. |
| `difficulty_raw` | int64 | Source-provided integer. |
### `doctor_annotated` — 5,250 rows, 15,151 human answers
| Field | Type | Notes |
|---|---|---|
| `subset_id` | int64 | Unique row id within this config. |
| `qa_id` | int64 / null | Joins to `qa.id`. Null for 10 rows whose question was removed from `qa` during cleaning. |
| `question`, `choices`, `answer`, `explanation`, `topic` | | Same meaning as in `qa`. |
| `annotations` | list of struct | One entry per doctor who answered this question. |
| `annotations[].rater_id` | int64 | Pseudonymous practitioner id (102 distinct). Stable across rows, so per-rater analysis is possible. |
| `annotations[].selected` | string | Letters the rater chose, sorted. |
| `annotations[].is_correct` | bool | Exact match against `answer` (multi-answer questions require the full set). |
| `annotations[].perceived_difficulty` | int64 | Rater's own 1–5 rating. |
| `annotations[].categories` | list of string | Rater's assignment into 10 TCM domains; empty list where the rater gave none. |
| `num_raters` | int64 | 3 for 4,801 rows, 2 for 299, 1 for 150. |
| `num_correct_raters` | int64 | Convenience count for majority-vote scoring. |
## Usage
```python
from datasets import load_dataset
qa = load_dataset("TechTCM/TCMQA", split="train") # 33,872
doc = load_dataset("TechTCM/TCMQA", "doctor_annotated", split="train") # 5,250
```
Rendering a prompt, skipping the padded fifth option:
```python
def render(row):
letters = "ABCDE"[: row["num_choices"]]
opts = "\n".join(f"{c}. {row['choices'][c]}" for c in letters)
return f"{row['question']}\n{opts}"
```
## Human baseline
Computed directly from the `doctor_annotated` config:
| Measure | Value |
|---|---|
| Practitioners | 102 |
| Total answers | 15,151 |
| Per-answer accuracy (all 15,151) | **62.10%** |
| 3-rater majority-vote accuracy (4,801 questions) | **64.94%** |
Both numbers are correct and measure different things. Use **62.10%** for "how well does
an individual practitioner do on an item", and **64.94%** when comparing a single model
prediction against a panel of three practitioners on the 4,801-question subset — the
latter is the appropriate baseline for most model-vs-human claims.
Doctors answered under exam-like conditions without reference material. Majority vote on
a 3-rater panel uses `num_correct_raters >= 2`.
## Credit
Built and maintained by **[TechTCM](https://techtcm.com/)**.
## License
Released under **CC BY-NC-SA 4.0**. Non-commercial use, attribution required, derivatives
under the same terms. See the provenance section regarding upstream material.
|