List your company

ResearchFormerly WizWand

LLM Safety Evaluation

Benchmarks

DatasetSOTA methodMetricTrendResultsLast updated

HARMAMP Benign instances

8.9Over-Refusal Rate

30

Jun 2, 2026

HARMAMP (malicious instances)

86.36Harm Score

30

Jun 2, 2026

Aymara Risk and Responsibility Matrix 10 Domains

Claude Haiku 3.5

86.2Mean Safety Score

20

May 1, 2026

9.88Ssafe Score

18

May 11, 2026

Unsafe and Jailbreak prompts reference and calibration models

meta-llama/Llama-3.1-8B-Instruct

0.283Unsafe Score

8

Jun 25, 2026

Goekdeniz-Guelmez/Josiefied-Qwen2.5-7B-Instruct-abliterated-v2

-0.272Unsafe Score

6

Jun 25, 2026

Safety Diagnostics LLaMA-3-8B (Unseen Anchor)

100Unseen Anchor Accuracy

4

Mar 20, 2026

Harmful Prompt Dataset

37.5Refusal Rate (Full)

3

Feb 26, 2026