| --- |
| license: cc-by-nc-sa-4.0 |
| language: |
| - en |
| size_categories: |
| - 10K<n<100K |
| tags: |
| - multimodal |
| - moral-reasoning |
| - vision-language |
| - benchmark |
| configs: |
| - config_name: default |
| data_files: |
| - split: test |
| path: test.parquet |
| --- |
| |
| # MM-SCALE — multimodal moral-acceptability evaluation set |
|
|
| **21,977 (image, scenario) pairs** across **8,444 DALL·E-generated images**, each |
| labeled with a single `mean_rating ∈ [1, 5]` for moral acceptability and a |
| single `modality_label ∈ {text, image, both}` indicating which modality the |
| judgment hinges on. |
|
|
| This is a **simplified eval view** — one row per (image, scenario) pair with one |
| rating and one modality label. The underlying paired-annotation source (per- |
| annotator ratings + modality votes, mixing humans and a GPT-5.2 oracle) is |
| documented in the accompanying paper. |
|
|
| ## Schema (per row) |
|
|
| | field | type | description | |
| |---|---|---| |
| | `image_id` | string | image identifier (matches the file in `images.tar.gz`) | |
| | `image_filename` | string | basename of the image, e.g. `00010.jpg` | |
| | `scenario_id` | string | unique scenario id, e.g. `00010.png_s01` | |
| | `target_setting` | string | original caption / situation description | |
| | `text` | string | the action being rated | |
| | `mean_rating` | float | mean of all annotator ratings, on the 1–5 scale (1=clearly unacceptable, 5=clearly acceptable) | |
| | `modality_label` | string | `"text"`, `"image"`, or `"both"` — which modality grounds the judgment | |
| | `n_annotations` | int | number of paired (rating, modality) annotations underlying the row (1, 2, or 3) | |
|
|
| ## How `mean_rating` is derived |
| |
| For each scenario, available annotations include (a) original human annotators |
| from MM-SCALE and (b) GPT-5.2 oracle annotations that filled missing slots up |
| to 3 paired (rating, modality_vote) entries. `mean_rating` is the arithmetic |
| mean of the `n_annotations` ratings. |
|
|
| ## How `modality_label` is derived |
| |
| 1. **Singleton or unanimous human votes** → use the human label. |
| 2. **Human votes disagree** → pool human + oracle votes and take the overall |
| majority. |
| 3. **Still tied** (rare) → fall back to the first human vote. |
| |
| Coverage of the resolution paths in this release: |
| |
| - 15,040 rows: singleton human vote |
| - 4,730 rows: ≥2 humans agreed |
| - 2,207 rows: humans disagreed; resolved with the oracle |
| |
| ## Distributions |
| |
| | `mean_rating` (rounded) | count | share | |
| |---|---|---| |
| | 1 | 1,118 | 5.1% | |
| | 2 | 3,038 | 13.8% | |
| | 3 | 4,893 | 22.3% | |
| | 4 | 8,341 | 38.0% | |
| | 5 | 4,587 | 20.9% | |
|
|
| | `modality_label` | count | share | |
| |---|---|---| |
| | image | 11,055 | 50.3% | |
| | both | 6,099 | 27.8% | |
| | text | 4,823 | 21.9% | |
|
|
| | `n_annotations` | count | |
| |---|---| |
| | 1 | 8,599 | |
| | 2 | 4,122 | |
| | 3 | 9,256 | |
|
|
| ## Images |
|
|
| Images ship as a separate archive `images.tar.gz` (full-resolution JPEG, q=90, |
| no downsizing) at this same repo. Reference each row's `image_filename` field |
| to locate the file under `images/<image_filename>` after extraction. |
|
|
| ```python |
| from huggingface_hub import hf_hub_download |
| hf_hub_download("mmscale/mmscale-data", "images.tar.gz", repo_type="dataset") |
| ``` |
|
|
| ## Loading |
|
|
| ```python |
| from datasets import load_dataset |
| ds = load_dataset("mmscale/mmscale-data") |
| print(ds["test"][0]) |
| # {'image_id': '00010.png', |
| # 'image_filename': '00010.jpg', |
| # 'scenario_id': '00010.png_s01', |
| # 'target_setting': '...', |
| # 'text': '...', |
| # 'mean_rating': 3.6667, |
| # 'modality_label': 'image', |
| # 'n_annotations': 3} |
| ``` |
|
|
| ## Caveats |
|
|
| - Images are synthetic (DALL·E); transfer to natural images is not validated |
| except on a small held-out SMID subset. |
| - English-only. |
| - 9,256 of 21,977 rows have `n_annotations == 3` (full paired coverage); the |
| rest (1 or 2 annotations) carry less-stable means. Filter on `n_annotations` |
| if your application needs fully-attested rows only. |
| - Oracle modality votes follow a stricter "decisiveness" reading than humans |
| appear to use; pooled-majority resolution mitigates but does not eliminate |
| this in the 2,207 tiebreak rows. |
|
|
| ## License |
|
|
| This dataset is released under the |
| **Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International |
| License** (CC-BY-NC-SA 4.0): |
| https://creativecommons.org/licenses/by-nc-sa/4.0/legalcode |
|
|
| The license is inherited from **Commonsense Norm Bank**, the upstream source |
| of the moral situations used to author scenarios in MM-SCALE. Per the upstream |
| terms: |
|
|
| > Commonsense Norm Bank is licensed under the Creative Commons |
| > Attribution-NonCommercial-ShareAlike 4.0 International License |
| > (https://creativecommons.org/licenses/by-nc-sa/4.0/legalcode). By requesting |
| > access to the dataset, you agree to follow the terms outlined in the |
| > license. |
|
|
| By using MM-SCALE you agree to the same terms (attribution, non-commercial |
| use, share-alike). |
|
|
| ### Accessing the original Commonsense Norm Bank |
|
|
| If you need the original Commonsense Norm Bank corpus (e.g. to reproduce |
| scenario authoring upstream of MM-SCALE), submit an access request here: |
|
|
| https://docs.google.com/forms/d/e/1FAIpQLSfk47SwWW4pUay4TMSFO7IaG2vxwpbUPI-TJIZwhwdBDNGqoQ/viewform |
|
|
| MM-SCALE itself does **not** redistribute the unmodified Commonsense Norm Bank |
| corpus; only the scenarios authored on top of those moral situations together |
| with their image pairings and annotations. |
|
|
| ## Citation |
|
|
| ```bibtex |
| @inproceedings{mmscale, |
| title={MM-SCALE: ...}, |
| author={...}, |
| year={2026} |
| } |
| ``` |
|
|
