FEAT Add SimpleSafetyTests dataset loader#1426
Conversation
romanlutz
force-pushed
the
romanlutz/add-simple-safety-tests-dataset
branch
from
7a32aee to
0b9827a
Compare
There was a problem hiding this comment.
Pull request overview
Adds a new remote seed dataset loader for HuggingFace dataset Bertievidgen/SimpleSafetyTests so it can be discovered and fetched via PyRIT’s SeedDatasetProvider mechanism.
Changes:
- Introduces
_SimpleSafetyTestsDatasetremote loader that fetches prompts from HuggingFace and converts them intoSeedPrompts. - Adds unit tests covering the new loader’s
fetch_datasetbehavior anddataset_name. - Updates dataset-loading documentation notebook output to include
simple_safety_testsin the available dataset list.
Reviewed changes
Copilot reviewed 4 out of 4 changed files in this pull request and generated 4 comments.
| File | Description |
|---|---|
pyrit/datasets/seed_datasets/remote/simple_safety_tests_dataset.py |
New HuggingFace-backed dataset loader for SimpleSafetyTests. |
pyrit/datasets/seed_datasets/remote/__init__.py |
Exposes the new loader so it registers/discovers with other remote providers. |
tests/unit/datasets/test_simple_safety_tests_dataset.py |
Adds unit tests for the new loader. |
doc/code/datasets/1_loading_datasets.ipynb |
Updates the “Loading Built-in Datasets” notebook to reflect the new dataset name, but also includes newly-executed metadata/outputs. |
| authors = ["Bertie Vidgen", "Nino Scherrer", "Hannah Rose Kirk", "Rebecca Qian", | ||
| "Anand Kannappan", "Scott Niekum", "Paul Röttger"] |
There was a problem hiding this comment.
The authors list formatting is inconsistent with the rest of the remote dataset loaders and makes it harder to read/maintain. Consider formatting it as a multi-line list with one author per line (as done in other loaders) to keep style consistent.
| authors = ["Bertie Vidgen", "Nino Scherrer", "Hannah Rose Kirk", "Rebecca Qian", | |
| "Anand Kannappan", "Scott Niekum", "Paul Röttger"] | |
| authors = [ | |
| "Bertie Vidgen", | |
| "Nino Scherrer", | |
| "Hannah Rose Kirk", | |
| "Rebecca Qian", | |
| "Anand Kannappan", | |
| "Scott Niekum", | |
| "Paul Röttger", | |
| ] |
| "cell_type": "code", | ||
| "execution_count": null, | ||
| "id": "1", | ||
| "metadata": {}, | ||
| "execution_count": 1, | ||
| "id": "1c3a2497", | ||
| "metadata": { | ||
| "execution": { | ||
| "iopub.execute_input": "2026-03-01T14:35:40.418153Z", | ||
| "iopub.status.busy": "2026-03-01T14:35:40.417953Z", | ||
| "iopub.status.idle": "2026-03-01T14:35:45.183664Z", | ||
| "shell.execute_reply": "2026-03-01T14:35:45.183155Z" | ||
| } | ||
| }, |
There was a problem hiding this comment.
This notebook now includes execution metadata (execution_count, iopub timestamps, etc.). Most notebooks in doc/code keep execution_count as null and omit per-run execution metadata to avoid noisy diffs and non-reproducible timestamps; consider clearing execution metadata before committing.
| "Found default environment files: ['C:\\\\Users\\\\romanlutz\\\\.pyrit\\\\.env', 'C:\\\\Users\\\\romanlutz\\\\.pyrit\\\\.env.local']\n", | ||
| "Loaded environment file: C:\\Users\\romanlutz\\.pyrit\\.env\n", | ||
| "Loaded environment file: C:\\Users\\romanlutz\\.pyrit\\.env.local\n" | ||
| ] | ||
| }, | ||
| { | ||
| "name": "stderr", | ||
| "output_type": "stream", | ||
| "text": [ | ||
| "C:\\Users\\anasundar\\AppData\\Local\\Temp\\ipykernel_49000\\4021500943.py:10: DeprecationWarning: is_objective parameter is deprecated since 0.13.0. Use seed_type='objective' instead.\n", | ||
| "C:\\Users\\romanlutz\\AppData\\Local\\Temp\\ipykernel_10540\\4021500943.py:10: DeprecationWarning: is_objective parameter is deprecated since 0.13.0. Use seed_type='objective' instead.\n", | ||
| " memory.get_seeds(harm_categories=[\"illegal\"], is_objective=True)\n" | ||
| ] | ||
| }, |
There was a problem hiding this comment.
The committed outputs include local machine paths/usernames (e.g., C:\\Users\\...) from environment loading. This can leak personal/workstation details into the repo; please clear these outputs (or replace them with sanitized, non-user-specific examples) before committing the notebook.
| source="https://huggingface.co/datasets/Bertievidgen/SimpleSafetyTests", | ||
| authors=authors, | ||
| groups=["Patronus AI", "University of Oxford", "Bocconi University"], | ||
| metadata={"category": item.get("category", "")}, |
There was a problem hiding this comment.
groups is set to [item.get("category", "")], which will produce a list containing an empty string when the category is missing/blank. Other dataset loaders avoid this by using an empty list when the category is falsy (e.g., groups=[category] if category else []), which prevents polluting group filters/queries with "".
| metadata={"category": item.get("category", "")}, | |
| metadata={"category": item.get("category")}, |
Add remote dataset loader for SimpleSafetyTests (Bertievidgen/SimpleSafetyTests), a lightweight diagnostic set of 100 critical safety test prompts for quickly evaluating the most basic safety properties of LLMs. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
romanlutz
force-pushed
the
romanlutz/add-simple-safety-tests-dataset
branch
from
0b9827a to
1efaa09
Compare
Add remote dataset loader for SimpleSafetyTests (Bertievidgen/SimpleSafetyTests), a lightweight diagnostic set of 100 critical safety test prompts for quickly evaluating the most basic safety properties of LLMs.