README.md

---
pretty_name: EndoDiffVQA
license: other
license_name: upstream-non-commercial-research
license_link: LICENSE
language:
- en
task_categories:
- visual-question-answering
- video-text-to-text
tags:
- surgery
- surgical-video
- endoscopy
- laparoscopy
- comparative-reasoning
- multiple-choice
- benchmark
- medical
size_categories:
- 10K<n<100K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train.parquet
  - split: validation
    path: data/val.parquet
  - split: test
    path: data/test.parquet
extra_gated_heading: Access to EndoDiffVQA
extra_gated_prompt: >-
  EndoDiffVQA contains surgical video derived from CholecT50, MultiBypass140 and CholeScore.
  Access is granted for non-commercial academic research only, and is subject to the licence of
  each upstream release (CholecT50 and MultiBypass140 are CC BY-NC-SA 4.0). You remain responsible
  for complying with those licences and for obtaining any further permission your use requires.
extra_gated_fields:
  Full name: text
  Institutional email: text
  Affiliation: text
  Intended use: text
  I will use this dataset for non-commercial academic research only: checkbox
  I will comply with the upstream licences of CholecT50, MultiBypass140 and CholeScore: checkbox
  I will not attempt to identify any patient or clinician appearing in the video: checkbox
  I will not redistribute the video files: checkbox
extra_gated_button_content: Request access
---

# EndoDiffVQA

**Comparative surgical video question answering.** Every item shows **two** surgical video clips and
asks one four-option multiple-choice question that cannot be answered from either clip alone — the
answer is a relation *between* the videos (same / different / present in only one / present in neither).

10,939 items · **7,289 train / 1,350 validation / 2,300 test** ·
1,762 clips (33.2 GB) · 4 sources · 4 categories ·
20 attributes · 14 question templates

This release is a **pin**: a frozen snapshot of the template-MCQ dataset as built on
2026-09-27, from commit `8cdde6465a6a`. Digests of every input file are in
`metadata/pin.json`.

## Sources

| source | procedure | items | train / val / test | categories | clips | video |
|---|---|---:|---|---|---:|---:|
| `cholect50_caption` | laparoscopic cholecystectomy | 6,190 | 4,396 / 632 / 1,162 | Action, Anatomy, Tool | 1,224 | 1.44 GB |
| `multibypass_caption` | laparoscopic Roux-en-Y gastric bypass | 2,060 | 1,104 / 468 / 488 | Action, Anatomy, Tool | 218 | 1.15 GB |
| `chole_score` | laparoscopic cholecystectomy | 1,684 | 1,089 / 175 / 420 | Skill | 201 | 22.82 GB |
| `multibypass_skill` | laparoscopic Roux-en-Y gastric bypass | 1,005 | 700 / 75 / 230 | Skill | 119 | 7.78 GB |

`cholect50_caption` and `multibypass_caption` carry the caption-derived categories (Anatomy, Tool,
Action); `chole_score` and `multibypass_skill` carry Skill. **No item ever pairs clips from two
different sources**, and no item pairs a CholeScore video with a MultiBypass one.

## Categories and attributes

| category | items | attributes | |
|---|---:|---:|---|
| Anatomy | 3,326 | 4 | `pathology`, `presence`, `color`, `position` |
| Tool | 3,241 | 4 | `count`, `presence`, `engagement`, `entry side` |
| Skill | 2,689 | 9 | `cystic artery dissection`, `fossa dissection`, `cystic duct dissection`, `Calot exposure`, `bimanual_dexterity`, `efficiency` … |
| Action | 1,683 | 3 | `rationale`, `interaction`, `target` |

Full taxonomy with per-split counts: `metadata/taxonomy.json`.

## Question construction

Questions are **template-generated**, not model-written: a template renders a stem plus four options
over facts read off the source annotation, so the ground truth is traceable to the annotation rather
than to a language model. `metadata/templates.json` lists every template with its stem and the
relations it realises.

Two option families:

- **Family E** — enumerated lattice — the four options are drawn from a fixed cell set, identical whatever the truth, so the shown set leaks nothing about the answer
- **Family V** — value-bearing — options quote concrete observed values (a colour, a count, a rubric level); distractor values come from the same source's value pool

Each distractor carries a code in `distractor_codes` (aligned with `options`):
- `correct` — the true option
- `D1` — SWAP — the two videos' values or the direction of the gap are exchanged
- `D2` — COLLAPSE — the wrong same/different verdict
- `D3` — MAGNITUDE — right verdict, wrong size of the gap
- `D4` — SUBSTITUTE — a plausible value drawn from the source's pool for this subject

**Balance.** Answer letters are a: 25.0% / b: 25.0% / c: 25.0% / d: 25.0%, so the
majority-letter baseline is 25.1%. The yes/no verdict
splits No 56.4% / Yes 43.6%; a text-only model
that always answers the majority verdict gets 56.4%. The skew is
uneven by family — E: No 50.7% / Yes 49.3%, V: No 63.7% / Yes 36.3% —
so report Family E and Family V separately. Family E items show a fixed option set independent of the
truth, so the option *text* does not reveal the answer.

## Relations

`relation` is the coarse answer type, `cell` the fine option slot it was drawn from
(`metadata/templates.json` → `_legend`):

- `SAME` (4,715) — both videos carry the same value, or the same standard
- `DIFFERENT` (3,774) — the two videos differ
- `ONLY` (2,054) — the subject is present in only one of the two videos
- `NEITHER` (396) — the subject is present in neither video

## Splits

Splits are cut to be **disjoint in the unit that carries the label** — case for the Skill sources,
clip for the caption sources — so no video appearing in `train` reappears in `val` or `test`. The
realised overlap counts are in `metadata/stats.json` under `disjointness`.

## Fields

| field | type | meaning |
|---|---|---|
| `id` | str | `{source}-template_mc-n2-{split}{pair}-{index}` |
| `split`, `source`, `qa_type`, `n_compare` | str, str, str, int | `split` is `train`/`val`/`test` (the `datasets` split is named `validation`); `qa_type` is `template_mc`; `n_compare` is 2 throughout |
| `category`, `attribute` | str | Anatomy / Tool / Action / Skill, and the compared attribute |
| `question` | str | stem with the four options inlined — feed this verbatim |
| `question_stem` | str | the stem alone |
| `options` | list[str] | the four option texts, in order a, b, c, d |
| `option_a` … `option_d` | str | the same, as flat columns |
| `answer_letter`, `answer`, `answer_text` | str | `c`; `(c) Yes — …`; `Yes — …` |
| `verdict` | str | `Yes` / `No` — the polarity of the correct option |
| `verdict_shared` | bool | whether the verdict is the same for every item of this group |
| `videos` | list[str] | release-relative clip paths, in order |
| `video1`, `video2` | str | the same, as flat columns |
| `template_id`, `family`, `relation`, `cell` | str | how the item was built |
| `distractor_codes` | list[str] | per-option code, aligned with `options` |
| `subject_keys` | list[str] | what the question is about (entity, tool, criterion, or [step, domain]) |
| `pair_id`, `qa_index` | str | position within the source build |
| `phase` | str \| null | surgical phase of the clip pair (caption sources) |
| `value_a`, `value_b` | str \| null | the compared values (Family V, caption sources) |
| `clip1`, `clip2` | int \| null | source clip ids (caption sources) |
| `distractor_pool` | str \| null | JSON: value → where the distractor was drawn from |
| `criterion`, `step` | str \| null | rated criterion and surgical step (Skill sources) |
| `score_a`, `score_b`, `delta` | int \| null | the two ratings and their difference (Skill sources) |
| `levels_shown` | list[int] | rubric levels quoted by the options (S2 only; else empty) |
| `unit1`, `unit2`, `case1`, `case2` | str \| null | rated unit and case id (Skill sources) |

## Usage

```python
from datasets import load_dataset
from huggingface_hub import snapshot_download

ds = load_dataset("ethanshili/EndoDiffVQA", split="test")          # metadata only, ~MBs
root = snapshot_download("ethanshili/EndoDiffVQA", repo_type="dataset")  # + video, 33 GB

item = ds[0]
print(item["question"])
paths = [f"{root}/{p}" for p in item["videos"]]           # the two clips, in order
```

Grading is exact-match on `answer_letter`. Report accuracy broken down by `source`, `category`,
`attribute` and `relation` — aggregate accuracy hides that the four sources ask different questions.

## Known limitations

- **Resolution is a recording-setup fingerprint.** Clip shapes are
  854x480 (×1,442), 1920x1080 (×205), 600x480 (×115).
  11.3% of items pair two clips of different size, and the share is uneven across
  splits (train 10.2%, val 0.0%, test 21.2%), so a model
  keying on frame geometry is not equally visible in validation and test. Normalisation to a common
  784×448 bound is planned and **not applied here**.
- **Option text of the CholeScore rubric levels (`S2`) is a DRAFT layer.** Levels 1/3/5 quote the
  upstream anchor verbatim; levels 2 and 4 are authored interpolations, and `S2` is restricted to
  1/3/5 in this build.
- **Templates are not equally supplied.** Per-template counts range widely (see
  `metadata/stats.json` → `split_x_template`); `C2` in particular is far below the others.
- **`n_compare` is 2 for every item.** The 3- and 4-video comparisons of the open-ended EndoDiffVQA
  splits are not part of this MCQ release.

## Licence and citation

Non-commercial academic research only. Each source keeps its upstream licence:

- **cholect50_caption** — CholecT50 (CAMMA, University of Strasbourg): CC BY-NC-SA 4.0 (registration form) — per the upstream project page  ⚠️ *not verified against an upstream LICENSE file — confirm before relying on it*
- **multibypass_caption** — MultiBypass140 (CAMMA — StrasBypass70 + BernBypass70): CC BY-NC-SA 4.0 (non-commercial research, no registration)
- **chole_score** — CholeScore (OSATS STS skill annotations over cholecystectomy video): unknown — confirm with the data provider before redistributing  ⚠️ *not verified against an upstream LICENSE file — confirm before relying on it*
- **multibypass_skill** — MultiBypass140 skill annotations (GOALS): CC BY-NC-SA 4.0 (non-commercial research, no registration)

Cite the upstream datasets alongside this release:

- Nwoye et al., Rendezvous: Attention mechanisms for the recognition of surgical action triplets in endoscopic videos, Medical Image Analysis, 2022.
- Lavanchy et al., Challenges in multi-centric generalization: phase and step recognition in Roux-en-Y gastric bypass surgery, IJCARS, 2024.
- Lavanchy et al., IJCARS, 2024 (skill annotations of the same release).