Feature request: speciesist bias harm definition and seed prompts
维护者通常 2 天内回复
评估
调研方向
Start by reading the existing harm definitions in pyrit/datasets/harm_definition/ and follow their format for a new speciesist-bias definition with the proposed 1–5 scale. Then inspect how local seed prompts are represented and loaded, and add paired prompts that differ by species. Done means the rubric evaluates inconsistent treatment and dismissiveness, without scoring dietary recommendations.
由索引模型根据 Issue 内容生成。
描述
Is your feature request related to a problem? Please describe.
Moved here from fairlearn/fairlearn#1625, where @romanlutz said this fits PyRIT's scope.
PyRIT can probe for representational harm against groups of people, but not for speciesist bias: models treating harm to some animals as acceptable or invisible while treating the same harm to other animals as wrong. This is now measured in the literature:
- Hagendorff, Bossert, Tse & Singer, "Speciesist bias in AI: how AI applications perpetuate discrimination and unfair outcomes against animals", AI and Ethics (2022), https://doi.org/10.1007/s43681-022-00199-9
- "Speciesism in AI: Evaluating Discrimination Against Animals in Large Language Models" (SpeciesismBench), https://arxiv.org/abs/2508.11534
- "What do Large Language Models Say About Animals? Investigating Risks of Animal Harm in Generated Text" (AHA benchmark), https://arxiv.org/abs/2503.04804
Describe the solution you'd like
A small first step that follows existing patterns:
- A harm definition in
pyrit/datasets/harm_definition/(for examplespeciesist_bias.yaml) with a 1–5 scale running from "weighs animal interests consistently" to "dismisses or hides serious harm to some species while recognising it for others". - A few local seed prompts that test for inconsistency, such as paired prompts that are identical except for the species (dog vs pig, cat vs chicken).
The rubric would score inconsistency and dismissiveness, not whether a response recommends a particular diet.
Describe alternatives you've considered, if relevant
Adding remote dataset loaders for the benchmarks above. I'd keep that for later, once licensing for each dataset is confirmed.
Additional context
I'm happy to open the PR. Would you prefer a new harm category, or adding this under representational?
- 主要语言
- Python
- 星标
- 4.7k
- 派生
- 980
- 平均合并
- 2 天 7 小时
- 30 天内合并 PR
- 446
环境准备
在浏览器里用你自己的 GitHub 账号启动这个项目的开发容器。
- 没有 Dockerfile 或 Docker Compose 文件
- 有 Pull Request 模板
- 没有贡献指南
从这里开始
- 先读完整个 Issue，再读项目的贡献指南。
- 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
- Fork 仓库，在一个分支上完成修改。
- 提交 Pull Request，并在描述里引用这个 Issue 编号。
microsoft/PyRIT 的其他 Issue
-
BUG: Garak and ML-VLSU loaders still load one example when max_examples is 0 or negative可能已有人在做 @ShivamSharma43 于 2 天前认领。 未关闭
难度 2/5 1-3 小时 新手友好度 78/100
维护者通常 2 天内回复
-
BUG: ComicJailbreak max_examples counts goals that render nothing, so templates=["code"] with max_examples=10 fails可能已有人在做 @ShivamSharma43 于 2 天前认领。 未关闭
难度 2/5 1-3 小时 新手友好度 76/100
维护者通常 2 天内回复
-
BUG: SeedDatasetFilter matches harm_categories, modalities and size case-sensitively, unlike tags可能已有人在做 @ShivamSharma43 于 2 天前认领。 未关闭
难度 2/5 1-3 小时 新手友好度 82/100
维护者通常 2 天内回复
-
Conversation drawer shows the previous attack's conversations after switching attacks可能已有人在做 @ManoharPaturi 于 2 天前认领。 未关闭
难度 2/5 1-3 小时 新手友好度 78/100
microsoft/PyRIT#3575 · 2 条评论 · 已指派 1 人 ·
维护者通常 2 天内回复
-
Conversation draft turns permanently dirty when applied conversions match no piece可能已有人在做 @ManoharPaturi 于 2 天前认领。 未关闭
难度 2/5 1-3 小时 新手友好度 83/100
microsoft/PyRIT#3573 · 已指派 1 人 ·
维护者通常 2 天内回复
相似的 Issue
-
难度 2/5 1-3 小时 新手友好度 72/100
ComposioHQ/awesome-claude-skills#2398 ·
维护者通常 2 天内回复
-
bug converter
难度 2/5 1-3 小时 新手友好度 62/100
grafana/sigma-rule-deployment#435 ·
维护者通常 1 天内回复
-
bug
难度 2/5 1-3 小时 新手友好度 73/100
datamodel-code-generator/datamodel-code-generator#5589 ·
维护者通常 1 天内回复
-
Shell: serve the home app's Cookie Policy on a hard reload可能已有人在做 @2001anshkaushik 今天认领。 未关闭feature module:ai
难度 2/5 1-3 小时 新手友好度 88/100
rocketride-org/rocketride-server#2899 · 1 条评论 ·
维护者通常 2 天内回复
-
GEFSv12
难度 2/5 1-3 小时 新手友好度 70/100