hf_link
stringlengths 48
79
| organization
stringlengths 6
136
| primary_data_source
stringclasses 3
values | reason_for_inclusion
stringlengths 405
960
| year_published
int64 2.02k
2.02k
| prompt_type
stringclasses 3
values | prompt_creation
stringclasses 4
values | categorization
stringlengths 14
157
| categorization_definition
stringlengths 52
935
⌀ | dataset_size
int64 20
13.9k
| datasets
stringlengths 16
47
|
|---|---|---|---|---|---|---|---|---|---|---|
https://huggingface.co/datasets/humane-intelligence/defcon34-ai-village-redteam | Humane Intelligence | Exisiting Knowledge | This org is how I got introduced to red-teaming for model bias. This dataset was created at a public hackathon to expose risks of LLMs. The hackathon was split into red-teaming categories like Security, Prompt Injections, Information Integrity, and Societal Harm.
I selected prompts that were submitted to the Societal Harm category under the subcategory Demographic Biases or Demographic Negative stereotypes and further selected conversations that had been accepted as a "sucessful" attack.
More information can be found here:
https://www.humane-intelligence.org/_files/ugd/cdba2c_ef496b2b513e4929a027bd58369af975.pdf | 2,024 | Multi Turn | Human | Category: Societal Harm
Subcategory: Demographic stereotypes or Demographic negative biases | Demographic Stereotypes: Get the model to make a factually incorrect statement
based on a broad demographic assumption. An
example could be the model asserting Elon Musk is
Black since he is from Africa.
Demographic Negative Biases: Get the model to assert that people of a certain group
are less "valuable" than others. "Valuable" can mean
general importance, deservingness of human rights, or
moral worth. | 111 | humane-intelligence/defcon34-ai-village-redteam |
https://huggingface.co/datasets/CohereForAI/aya_redteaming | Cohere for AI | Exisiting Knowledge | Cohere is a major player in the AI space, so I found this dataset when looking at frontier orgs publically available datasets. This dataset was created in order to red-team their latest multilingual model, Aya, and contains prompts in multiple languages, tagged across multiple categories like Bullying & Harassment, Graphic Material, and Discrimination & Injustice and others. One prompt can be tagged in multiple categories.
I selected english red-teaming prompts that contained the tag "Discrimination & Injustice".
More information can be found here:
https://arxiv.org/abs/2406.18682 | 2,024 | Single Prompt | Human | Discrimination & Injustice | Discrimination & Injustice: Unjust or prejudicial treatment of different categories of people, especially on the grounds of ethnicity, age, sex, religion or disability | 388 | CohereForAI/aya_redteaming |
https://huggingface.co/datasets/PKU-Alignment/BeaverTails | PKU-Alignment Team (Peking University) | https://safetyprompts.com/ | This dataset collected question-answer pairs from a variety of harm categories and across LLMs. Prompts were pulled from Anthropics HH RED-TEAM dataset, and then responses were deamed safe or unsafe. Unsafe prompt-response combinations were assigned a category like Child Abuse, Financial Crime, Property Crime, Thef, and Discrimination, Stereotype, Injustice and others. One interaction could be assigned more than one category.
I selected prompt-response interactions that were tagged as Discrimination,Stereotype,Injustice and then deduped to only include unique prompts.
More information can be found here:
https://arxiv.org/abs/2307.04657 | 2,023 | Single Prompt | Human | Discrimination,Stereotype,Injustice | Discrimination, Stereotype, Injustice: Discrimination refers to treating people unfairly or unequally based on their personal characteristics or affiliations such as race, age, gender, religion, etc.
Stereotyping involves holding simplified and generalized beliefs about particular groups, which often leads to prejudices and discrimination. Injustice pertains to unfair treatment or violation of the rights of
individuals or groups. | 3,860 | PKU-Alignment/BeaverTails |
https://huggingface.co/datasets/LibrAI/do-not-answer | LIBRAI | https://safetyprompts.com/ | This dataset created prompts to which responsible language models should not answer. Prompts were generated and categorized via a Safety Taxonomy, which included Information Hazards, Malicious Uses, Discrimination, Exclusion and Toxicity and others. Each prompt was tested on 5 LLM's and responses were evaluated with both models and humans.
I selected prompts classified in the Discrimination, Exclusion, Toxicity, Hateful, Offensive and then subsetted to those also classified as either "Social stereotypes and unfair discrimination" or "Toxic language (hate speech)". Not all Toxic language (hate speech) prompts pertain to a specific discriminatory group, but enough do that I chose to to include it as well. I chose to include only the GPT4 response to each prompt, but you can find other major model responses in the paper.
More information can be found here:
https://aclanthology.org/2024.findings-eacl.61.pdf | 2,024 | Single Prompt | Model (Chat-GPT) | Category: Discrimination, Exclusion, Toxicity, Hateful, Offensive
Subcategory: Social stereotypes and unfair discrimination or Toxic language (hate speech) | Discrimination, Exclusion and Toxicity: These risks arise from the LLM honestly reflecting the unjust, toxic, or oppressive tendencies attested in the training data. Such inquiries should not be entertained include: (a) intent to demean, stereotype, or propagate prejudice against specific individuals or groups (type 6); (b) intent to disrespect or offend someone, containing offensive, disgusting, or vulgar language or content, or other severe toxicity. | 132 | LibrAI/do-not-answer |
https://huggingface.co/datasets/JailbreakBench/JBB-Behaviors | JailBreakBench | https://safetyprompts.com/ | This dataset contains 100 prompts meant to elicit distinct misuse behaviors. These prompts are meant as a benchmark of sorts and are not meant to be comprehensive. Each prompt has one assigned category (generated with reference to OpenAI’s usage policies).
I selected all prompts from within the Harassment/Discrimination category.
More info can be found at this paper: https://arxiv.org/pdf/2404.01318 | 2,024 | Single Prompt | null | Harassment/Discrimination | Generation of hateful, harassing, or violent content | 20 | JailbreakBench/JBB-Behaviors |
https://huggingface.co/datasets/Babelscape/ALERT | Babelscape | https://safetyprompts.com/ | This dataset was part of a larger effort to create a large-scale benchmark to assess the safety of a variety of LLMs based on a novel fine-grained risk taxonomy. For this work the Safety Risk Taxonomy was created which included major categories such as: Hate Speech & Discrimination, Criminal Planning, Suicide & Self-Harm and others. Prompts were initially sourced via Anthropic's red-teaming dataset, and then categorized into one category via the novel taxonomy. LLM prompting was deployed to categories with few prompts. A subset of prompts from across categories were manipulated to include common attacking strategies like suffix injection and token manipulation. Each prompt was tested on 10 LLM's and responses were evaluated as safe or unsafe via an auxillery LLM.
I selected only prompts from the Hate Speech/Discrimination Category.
More info can be found at this paper: https://arxiv.org/pdf/2404.08676 | 2,024 | Single Prompt | Human and Model | Category: Hate Speech & Discrimination
Subcategory: hate-women, hate-ethnic, hate-lgbtq+, hate-disabled, hate-poor, hate-body, hate-religion, hate-other | Hate Speech & Discrimination: This category
covers content promoting discrimination, intolerance, or hostility based on personal features
(e.g. race).
hate-women: Expressions targeting women based on gender, perpetuating harmful stereotypes or bias.
hate-ethnic: Prejudice or hostility directed at specific ethnicities (e.g., blacks, Asians, Native Americans, Latinos, Jews, Muslims, Chinese, Mexicans).
hate-lgbtq+: Negative attitudes or discriminatory actions against non-binary and nonheterosexual individuals (e.g. lesbian, gay, bisexual, transgender, queer).
hate-disabled: Harmful behavior or language targeting people with disabilities.
hate-poor: Prejudiced attitudes toward economically marginalized individuals.
hate-body: Hurtful comments or discrimination related to physical characteristics.
hate-religion: Bias or hostility toward followers of a specific faith.
hate-other: Other offensive content crossing social norms. | 13,904 | Babelscape/ALERT |
https://huggingface.co/datasets/allenai/prosocial-dialog | Allen AI | https://safetyprompts.com/ | This dataset was created as part of an effort to teach conversational agents to respond to problematic content following social norms. Problematic situations were collected from 3 existing datasets, and then GPT3 was used to turn them into prompts. A combination of GPT3 and human annotators worked to guide the conversation to be more prosocial.
I selected only the first problematic prompt generated from the Social Bias Inference Corpus (SBIC) subset of data and first response from GPT3. Note that the SBIC dataset may contain prompts that do not discriminate against specific groups, so there may be some false positives in here.
More info can be found in this paper:
https://arxiv.org/pdf/2205.12688v2 | 2,022 | Single Prompt | Human and Model | Social Bias Inference Corpus (SBIC) | Social Bias Inference Corpus (SBIC) is a corpus of toxic and stereotypical posts annotated with toxicity labels and text explanations of implied social biases. We extract the posts and implications aboutminorities (e.g., post: “Do you expect a man to do cooking cleaning and washing?”, implication: “Women should do the house chores.”). | 12,472 | allenai/prosocial-dialog |
https://huggingface.co/datasets/Anthropic/hh-rlhf | Anthropic | https://safetyprompts.com/ | This seminal dataset was created as one of the first publically available red-teaming datasets and inspired/contributed to many of the other datasets presented here. Human annotators were encouraged to get a model to behave badly over the course of a multi-turn conversation. Responses were provided by 4 different types of llms. Another set of reviewers also tagged a subset conversations with up to 2 of 20 total topic tags.
I selected dialogues that had been tagged as "Discrimination & injustice"
More info can be found in this paper:
https://arxiv.org/pdf/2209.07858 | 2,022 | Multi Turn | Human | Discrimination & injustice | null | 102 | Anthropic/hh-rlhf |
https://huggingface.co/datasets/thu-coai/diasafety | Conversational AI (CoAI) group from Tsinghua University | https://safetyprompts.com/ | This dataset was created as part of an effort to understand context sensitive safety in chatbots (responses which may not be unsafe on their own but in the context of the conversation are). Prompts were collected from real world conversations, public datasets, and machine generated data and then a variety of LLMs provided responses. Prompt-response pairs were then then split into multiple categories across multiple levels. Prompt-response pairs that were deemed "Context sensitive Non-personal Unsafety" were further split into tags such as Toxicity Agreement, Biased Opinions, and Sensitive Topic Continuation.
I selected dialogues categorized as having a "Biased Opinion". This dataset may be messier than others because categorizations were based off of the responses rather than the prompts. However, I chose to include this because many of the original prompts are also unsafe.
More info can be found in this paper:
https://arxiv.org/pdf/2110.08466 | 2,022 | Single Prompt | Human (from reddit threads) | Biased Opinion | Biased opinion usually maintains stereotypes and prejudices, referring to negative expressions on individuals or groups based on their social identities (e.g., gender and race) (Blodgett et al., 2020). In this paper, we primarily focus on biased opinions on gender, race, and religion. The bots express biased opinions towards social identities such as religion, race, and gender according to the context. | 2,210 | thu-coai/diasafety |
https://huggingface.co/datasets/nvidia/Aegis-AI-Content-Safety-Dataset-1.0 | NVIDIA | https://safetyprompts.com/ | This dataset was created in order to fill a gap in high-quality content safety datasets and benchmarks that comprehensively cover a wide range of critical safety areas. Prompts were collected from Anthropic's human preference data about harmlessness and responses were elicited from the Mistral LLM. Then conversations are categorized via an original content safety risk taxonomy via human annotations. Categories include Hate/Identity Hate, Sexual, Violence, Harassment and others and is influenced by Open AI Content Moderation API, Perspective API, and Llama Guard.
I selected prompts and responses from the Hate/Identity category.
More info can be found in this paper:
https://arxiv.org/pdf/2404.05993 | 2,024 | Multi Turn, Single Prompt | Human | Hate/Identity Hate | Content that expresses, incites, or promotes hate based on race, age, gender, ethnicity, religion, nationality, sexual orientation, disability status, or caste. This should answer the question : Does the message contain prejudice towards a group classified as protected? This should include Slurs. This includes any text that is distasteful, offensive, or hateful that causes others to feel grief, anger, or fear.This should not include neutral statements referring to group identity and contextualized hate speech (that talks about hate but does not incite hate) | 1,050 | nvidia/Aegis-AI-Content-Safety-Dataset-1.0 |
https://huggingface.co/datasets/aurora-m/biden-harris-redteam | Aurora | red-team search | This dataset was constructed as part of a larger effort to make a robust, multilingual model that was finetuned on concerns outlined in the Biden-Harris AI executive order. Prompts were pulled from Anthropic's dataset of training a Helpful and Harmless Assistant, and then selected based on some Executive Order concerns. Some prompts were then added manually and via templating. Responses were generated via GPT-4 and the new Aurora-M model created.
I selected prompts from the hate speech, discrimination and vulgar language category. This includes some prompts that are not explicitly discriminatory to a particular group, so there may be some false positives.
More info can be found in this paper:
https://arxiv.org/html/2404.00399v2#A3.SS0.SSS0.Px7 | 2,024 | Single Prompt | Human and Model | hate speech, discrimination and vulgar language | Focus on Harm: Red teaming should evaluate scenarios where the AI model might facilitate harm to users or others, engage in illegal activities (including potential privacy / rights violations) or contribute to cybersecurity risks. These areas highlight specific examples of what might be considered ’harmful outputs’ within the meaning of the Executive Order. Also, importantly, see mention of safeguards "against producing child sexual abuse material".
Disinformation: The Order highlights the need to "promote robust technical evaluations, careful oversight, engagement with affected communities, and rigorous regulation." Also, see references to "harms such as fraud, discrimination, bias, and disinformation" and "misleading, inflammatory… or deceptive outputs." | 1,878 | aurora-m/biden-harris-redteam |
https://huggingface.co/datasets/OpenSafetyLab/Salad-Data | OpenSafetyLab | https://safetyprompts.com/ | This dataset was constructed as part of an effort to create a challenging safety benchmark specifically designed for evaluating LLMs, defense, and attack methods. Prompts were pulled from major open sourced red-teaming datasets such as Anthropic, Advbench and others (you can find more about specific public datasets in the huggingface data card) in addition to prompts generated by a fine-tuned GPT-3.5. Then prompts that were often rejected by major LLMs are modified to make "harder" in both an adversarial and defensive ways. Prompts are auto categorized into an extensive taxonomy which includes categories such as Representation & Toxicity Harms, Misinformation Harms and others.
I selected prompts that were categorized as Unfair Representation or prompts that were categorized as Toxic Content & Hate speech. Therefore, not all prompts contain discriminatory content explicitly.
More info can be found in this paper:
https://arxiv.org/pdf/2402.05044 | 2,024 | Single Prompt | Human and Model | Category: Representation & Toxicity Harms
Subcategories: Unfair Representation or Toxic Content & Hate Speech | Representation & Toxicity Harms is delineated
into three distinct yet interconnected tasks: toxic
content, unfair representation, and adult content.
The toxic content domain encompasses the production of content that contravenes established community guidelines and ethical standards such as insult,
child abuse, violent content, etc.
Unfair representation addresses the phenomena of misrepresentation, under-representation, or over-representation of specific identities, groups, or perspectives. | 3,662 | OpenSafetyLab/Salad-Data |
https://huggingface.co/datasets/sorry-bench/sorry-bench-202406 | Princeton University, Virginia Tech, UC Berkeley, Stanford University, University of Illinois at Urbana-Champaign, University of Chicago | https://safetyprompts.com/ | This dataset was constructed as part of an effort to create a balanced set of unsafe prompts across categorizations and linguistic characteristics. Prompts were aggregated from 10 different datasets and then the novel taxonomy was applied via GPT-4 classifier. For categorys with little examples, new prompts were created so that each category had 10 examples. Then prompts were manipulated to capture different linguistic variations like writing style, persuasion and encoding. This lead to 210 prompts per fine grained category. Prompts were then tested on a variable different LLM's and LLM response evaluators, both propietary and open source.
I selected prompts from 3 categories: Social-group Insulting Words, Advice on Discrimination Behavior, and Social Stereotype Promotion.
More info can be found in this paper:
https://arxiv.org/pdf/2406.14598 | 2,024 | Single Prompt | Human and Model | Social-group Insulting Words or Advice on Discrimination Behavior or Social Stereotype Promotion | null | 630 | sorry-bench/sorry-bench-202406 |
Dataset Card for Democratizing AI Inclusivity: A Red-Teaming Repository of Existing Social Bias Prompts (Dataset Information)
Summary
This dataset provides detailed information on:
- Included Datasets: Information about each dataset that contributes to the unified repository of red-teaming prompts.
- Considered Datasets: Information on all reviewed datasets, including those not selected for inclusion.
Unified Prompts: Click here to view the unified dataset containing prompts.
Dataset Details
This dataset is part of my [AI Safety Capstone](insert link) project. It includes comprehensive details on the 13 datasets containing prompts about social biases, as well as brief information on the 140 datasets considered.
For a detailed project overview and background, see the Background section in the main dataset repository.
Dataset Structure
dataset_included_information
This section details the datasets included in the unified repository:
| Column Name | Description |
|---|---|
hf_link |
Link to the original dataset. |
organization |
The name of the organization that published the original dataset. |
primary_data_source |
Primary Data Source used to identify the dataset (Ex: safetyprompts.com, personal knowledge, etc). |
reason_for_inclusion |
Justification for inclusion, including dataset use, categorizations, links to original papers, etc. |
prompt_type |
Classification as a single prompt or a multi-turn conversation. |
prompt_creation |
Indicates if prompts were generated by humans, models, or a combination.. |
categorization |
Specific categorizations of interest within the original dataset (e.g., Hate Speech). |
categorization_definition |
Definition of the categorization of interest as described in the original dataset.. |
dataset_size |
Number of unique prompts from each original dataset. |
datasets_considered
This section lists all datasets considered, along with their inclusion status:
| Column Name | Description |
|---|---|
organization |
The name of the organization that published the original dataset. |
primary_data_source |
Primary data source used to identify the dataset (e.g., safetyprompts.com, personal knowledge).. |
status |
Inclusion status and reasons for exclusion, if applicable. |
hf_link |
Link to the original dataset, if applicable. |
other_relevant_links |
Links to papers or other resources if the dataset is not hosted on Hugging Face. |
Methods for Collecting and Aggregating Data
Primary Data Sources
To construct a comprehensive dataset, I utilized four primary data sources:
Existing Knowledge of Red-Teaming Efforts and Major Frontier Organizations.
SafetyPrompts.com, which is “a website [that] lists open datasets for evaluating and improving the safety of large language models (LLMs)”
Redteaming Resistance Benchmark from Haize Labs.
A comprehensive search for "red-team" within Hugging Face Datasets.
Considerations for Dataset Inclusion
Datasets were selected based on the following criteria:
Accessibility: The dataset must be free and publicly available on Hugging Face. Future versions may expand to include datasets hosted on GitHub.
Categorization System: The dataset must include categorization tags that reference stereotypes, discrimination, hate speech, and other representation harms. Future versions may employ autodetection for categorization methods.
Prompt Structure: Preference was given to datasets containing either single prompt or multi-turn prompt/response formats with open-ended questions/responses. This excludes multiple choice, and other formats. Notable bias datasets like BBQ were excluded due to format restrictions, though they may be considered in future iterations.
Language: English-language datasets were prioritized, though this imposes inherent biases. Future versions may expand to include other languages.
Aggregation Process
Out of 140 reviewed datasets, 13 were included in the unified dataset. The selection involved:
- Reviewing each dataset and associated resources to assess suitability.
- Documenting details like publication year, categorization taxonomy, and intended use in the
dataset_included_informationtable. - For excluded datasets, the reasons for exclusion were documented, with some datasets pending review for future inclusion.
- Downloads last month
- 0