Details

S-Eval (IS2Lab) family: a set of base risks and jailbreak attacks generated from them. It checks whether the model recognizes harmful intent in both the original and masked requests.

Harmful content filtering evaluation with focus on safety and commercial applicability: data is close to real moderation scenarios.

Toxicity benchmark in dialogues based on real user interactions with LLMs. The final label is formed as OR between toxicity and jailbreaking columns.

Universal high-quality benchmark based on WildGuardMix, translated into 17 languages. Tests toxicity, adversarial attacks, unsafe requests, and false positives (exaggerated safety).

ToxicityPrompts/RTP-LX: toxicity evaluation in multilingual environment considering cultural nuances and hidden forms of harm (microaggressions, bias) that often escape models.

A set of truly toxic prompts and a hard 1k subset (rejected by at least 3 of the largest SOTA models). Evaluates the balance between safety and usefulness, minimizing excessive refusals (over-refusal).

walledai/XSTest: diagnostics of model over-caution (false refusals on safe requests with trigger words) using contrastive pairs of safe and unsafe prompts.

Measures robustness to unambiguously malicious requests blocked by all market leaders. A response to such a prompt is considered a critical security failure (false negative). Translations are done by native speakers.

Checks whether the model remains safe and appropriate, evaluating robustness to harmful content generation on labeled prompt-response pairs and the balance between safety and usefulness.

Prompt-response pairs with human annotation plus request subsets. Tests robustness to a wide range of adversarial attacks: AutoDan, GCG, Human, Few-shot, PAIR, Persona.

Evaluates how accurately the model classifies human-translated real jailbreak requests in different languages, including low-resource ones (en, zh, it, vi, ar, ko, th, bn, sw, jv).

Bertievidgen/SimpleSafetyTests: a quick smoke-test / sanity-check consisting of concentrated malicious requests for rapid identification of critical security vulnerabilities.

walledai/CSRT: evaluates robustness to malicious requests disguised by mixing multiple languages in one phrase. Built on the basis of MultiJail.

CohereLabs/aya_redteaming: robustness to targeted attacks and correct handling of malicious requests in different cultural contexts, with separation of harm into "global" and "local".

The first multilingual safety benchmark: 14 types of risks in 10 languages. Tests robustness to unsafe requests and complex jailbreak attacks (role-playing, prompt leakage).

mmathys/openai-moderation-api-evaluation: prevention of content generation that violates safety policy, according to a fixed set of categories (violence, hate, sexual content, self-harm).

Real user requests and responses: malicious, safe, and adversarially-safe splits. Summary metrics FNR/FPR/F1 for evaluating overall robustness on native Russian data.

Augmented (complicated) versions of malicious and safe requests. Testing robustness to aggravated attack variants and false positives on complicated safe examples.

Dataset for evaluating robustness to prompt injections: examples of attacks aimed at hijacking model control, in Russian and English.