| """ |
| Text content for the GuardBench Leaderboard. |
| """ |
|
|
| TITLE = """ |
| <div style="text-align: center; margin-bottom: 1rem"> |
| <h1>GuardBench Leaderboard</h1> |
| </div> |
| """ |
|
|
| INTRODUCTION_TEXT = """ |
| ## Introduction |
| |
| GuardBench is a comprehensive benchmark for evaluating the safety guardrails of large language models (LLMs). |
| This leaderboard tracks model performance across various safety categories, including harmful content detection, |
| jailbreak resistance, and more. |
| |
| Models are evaluated on their ability to properly refuse harmful requests and detect problematic content |
| across multiple categories and test scenarios. |
| """ |
|
|
| LLM_BENCHMARKS_TEXT = """ |
| ## GuardBench Evaluation Methodology |
| |
| GuardBench evaluates models across multiple safety categories: |
| |
| - **Criminal, Violent, and Terrorist Activity** |
| - **Manipulation, Deception, and Misinformation** |
| - **Creative Content Involving Illicit Themes** |
| - **Sexual Content and Violence** |
| - **Political Corruption and Legal Evasion** |
| - **Labor Exploitation and Human Trafficking** |
| - **Environmental and Industrial Harm** |
| - **Animal Cruelty and Exploitation** |
| - **Self-Harm and Suicidal Ideation** |
| - **Safe Prompts** |
| |
| Each category is tested with: |
| - **Default Prompts**: Standard harmful requests |
| - **Jailbreaked Prompts**: Adversarial attempts to bypass safety measures |
| - **Default Answers**: Evaluation of harmful content in responses |
| - **Jailbreaked Answers**: Evaluation of harmful content in responses to jailbreak attempts |
| |
| Metrics include: |
| - **F1 Binary**: Harmonic mean of precision and recall |
| - **Recall Binary**: Proportion of actual harmful content correctly identified |
| - **Precision Binary**: Proportion of identified harmful content that was actually harmful |
| - **Error Ratio**: Rate of errors during evaluation |
| - **Average Runtime**: Average processing time in milliseconds |
| """ |
|
|
| EVALUATION_QUEUE_TEXT = """ |
| ## Submission Process |
| |
| To submit your model results to the GuardBench leaderboard: |
| |
| 1. Evaluate your model using the [GuardBench framework](https://github.com/huggingface/guard-bench) |
| 2. Format your results as a JSONL file according to our schema |
| 3. Submit your results using the submission form with your authorized token |
| |
| Results will be processed and added to the leaderboard once validated. |
| """ |
|
|
| CITATION_BUTTON_LABEL = "Cite GuardBench" |
|
|
| CITATION_BUTTON_TEXT = """ |
| @misc{guardbench2023, |
| author = {GuardBench Team}, |
| title = {GuardBench: Comprehensive Benchmark for LLM Safety Guardrails}, |
| year = {2023}, |
| publisher = {GitHub}, |
| journal = {GitHub repository}, |
| howpublished = {\\url{https://github.com/huggingface/guard-bench}} |
| } |
| """ |
|
|
