Download SOURCES.md from Alibaba-VELLDEPTH/MMA-SafetyBench: direct link, hf CLI and curl.
- Browser
- Download file 3.54 kB
-
https://huggingface.co/datasets/Alibaba-VELLDEPTH/MMA-SafetyBench/resolve/main/SOURCES.md
- Command line
-
hf download hf://datasets/Alibaba-VELLDEPTH/MMA-SafetyBench/SOURCES.md
-
curl -L -o SOURCES.md https://huggingface.co/datasets/Alibaba-VELLDEPTH/MMA-SafetyBench/resolve/main/SOURCES.md
Source attribution and redistribution review
Review date: 2026-10-01. Public availability alone is not treated as permission to redistribute all assets. Repository software licenses are not automatically assigned to screenshots, documents or videos.
Author confirmation, 2026-10-02: the dataset author confirmed that the video and image assets may be used and published. This resolves the request for author confirmation; the table below retains the independently documented provenance and gaps in source-version/license records. Third-party assets are not relabeled as author-original material. No additional permission claim is inferred from the confirmation.
| Component | Source | Status |
|---|---|---|
| Web | VisualWebArena | Source identified in manuscript; asset-level terms and release version still to be recorded. |
| Desktop GUI | OSWorld | Source identified in manuscript; asset-level terms and release version still to be recorded. |
| Mobile | OS-Sentinel, MobileRisk | OS-Sentinel code repository displays Apache-2.0; applicability to all distributed MobileRisk screenshots and associated third-party content is not established here. |
| Document: receipts | CORD, official CORD v2 | Authors confirmed CORD as the source. Official README declares CC BY 4.0. Twenty truncated base images were restored from the v2 test release using unique exact matches of intact pixels and dimensions. This identifies the recovery source, not the version originally used for all experiments. |
| Document: resumes | Resume Dataset PDF | Source cited in manuscript; dataset-page license and provenance not yet verified. |
| Video | LVBench | Source identified in manuscript; underlying video redistribution rights and source version still to be recorded. |
Modifications
MMA-SafetyBench adds adversarial visual content and associated task/evaluation annotations as described in the paper. The current candidate additionally localizes signed Mobile image references and removes identified hard-coded API keys from scripts. These changes do not relicense underlying third-party assets.
CORD citation
Recovery source: data/test-00000-of-00001-9c204eb3f4e11791.parquet from naver-clova-ix/cord-v2, SHA-256 51c65f1788faff392abe2a0b55b023eb23e9be551c509138eaa3a832514224e7. Original test IDs and local receipt filename numbers match for all 20 restored images. Subsequently, injected images 038 and 059 were reconstructed with the author-supplied generator and recorded new random seeds, as documented in the README. Experimental measurements were not changed or revalidated.
Seunghyun Park, Seung Shin, Bado Lee, Junyeop Lee, Jaeheung Surh, Minjoon Seo, and Hwalsuk Lee. 2019. CORD: A Consolidated Receipt Dataset for Post-OCR Parsing. Document Intelligence Workshop at NeurIPS. Paper.
CORD's official README records removal of some published annotation categories for legal reasons. This is not represented here as a guarantee that every image is free of personal information. Resumes and receipts are separate source collections; RVL-CDIP is not the confirmed receipt source for this package.