TIER (Qwen2.5-7B-Instruct Judge): leaderboard
TIER (Trinh-Thi et al., 2026) probes LLM safety with 1,184 prompts over sexual content, violence, illegal activity and self-harm, each risk area at four levels of threat implicitness from explicit requests to jailbreaks. Qwen2.5-7B-Instruct labels each greedy response on a six-label behavior scale; the board is the attack success rate, the percent of prompts answered with harmful compliance averaged over the four levels, and lower is better.
Metric: Overall attack success rate (%; mean of the four threat levels over 1,184 prompts in four risk areas; share of prompts whose response Qwen2.5-7B-Instruct labels harmful compliance (label 5 of a six-label behavior scale), greedy decoding with the default chat configuration). Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | gpt-j-6B | 23.9 | #9941 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=tier-qwen2-5-7b-instruct-judge · How It Works · Data refreshed daily, snapshot 2026-11-12.