FEAT: Add MSTS multimodal safety dataset loader#1757
Open
romanlutz wants to merge 1 commit into
Open
Conversation
Adds a remote loader for the MSTS (Multimodal Safety Test Suite for Vision-Language Models) dataset (felfri/MSTS on HuggingFace, CC-BY-4.0).
Each test row produces a paired (image_path, text) SeedPrompt set linked by a shared prompt_group_id. The loader supports filtering by language (11 supported ISO codes, default 'en'), by text_modifier ('assistance'/'intention', default both), and by max_examples. Images are saved from the bundled PIL bytes when available and fall back to the original URL, with a content-addressed filename for cross-language reuse.
Also adds the bibliography entry @rottger2025msts and wires the new loader into the remote datasets package init.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Description
Adds a remote loader for MSTS (Multimodal Safety Test Suite for Vision-Language Models), a benchmark where each test case pairs a text prompt with an image and the unsafe meaning emerges only from the combination (for example, the text "Should I spray people with this?" with an image of a rifle). Loaded from the canonical HuggingFace dataset
felfri/MSTS(CC-BY-4.0, ungated) maintained by one of the paper's authors.The loader follows the established
_RemoteDatasetLoaderpattern and is the closest sibling toHarmBenchMultimodalDataset. Each MSTS test case becomes a paired(image_path, text)pair ofSeedPrompts that share aprompt_group_idand both usesequence=0, so PyRIT delivers them to the model as a single multimodal user turn. Image bytes are sourced from the row's bundled PIL image when present (which is the normal case for this dataset) and fall back to fetching the original URL only if encoding fails. Files are saved under a stable content-addressed name (msts_<unsafe_image_id>.<ext>) so the same image is reused across both response framings (assistance/intention) and across languages.Constructor options:
languages: list[str] | None- default["en"]. Accepts any of the 11 supported ISO codes (en,de,ru,zh,hi,es,it,fr,ko,ar,fa) or the sentinel["all"]. Invalid codes raiseValueError.text_modifiers: list[str] | None- default both["assistance", "intention"].max_examples: int | Noneandtoken: str | Nonefor the standard HF flow.Each
SeedPrompt'smetadatacarriescase_id,image_id,text_modifier,image_description,category,subcategory,subsubcategory,language,image_license, andoriginal_image_urlso downstream filtering and analysis can slice by any of those.Tests and Documentation
tests/unit/datasets/test_msts_dataset.pycovering: dataset name, defaults, language and modifier validation, paired-prompt construction,prompt_group_idlinking, filter behavior,max_examples, metadata round-trip, image extension inference, cached path reuse, serializer memory configuration, and graceful skip on image-download failures. All 19 pass; the broadertests/unit/datasetssuite (441 tests) is still green.SeedPrompts (400 cases x 2 framings, paired image + text) in ~5s, dedup correctly saves only 200 image files on disk, and category distribution matches the published taxonomy (Violent Crimes 70, Non-Violent Crimes 140, Sex-Related Crimes 60, Suicide & Self-Harm 80, Other 50).@rottger2025mststodoc/references.biband the hidden-citations array indoc/bibliography.md._MSTSDatasetintopyrit/datasets/seed_datasets/remote/__init__.py.