Visualizations

The filters below control all charts in sync; “Pin” locks a model for comparison.

Models

Groups / benchmarks

Language
5 models · 19 groups Export CSV
Pinned for comparison Nothing pinned, press Pin in the Models list.

Group radar

scores of selected models by group · higher = better
0.250.500.751.00
S-Eval AEGIS 2.0 ToxicChat PolyGuard RTP-LX OR-Bench XSTest StrongReject++ BeaverTails HarmBench MultiJail SimpleSafetyTests CSRT Aya Red Teaming XSafety OpenAI Moderation Robustness Test (real) Robustness Test (robust) Prompt injection
YuFeng-XGuard-Reason-8B 8B Sing-Guard-2b 2B HiveTraceGuard-Pro 0.6B Shieldstral-1.0-3B 3B OpenGuardrails-Text-2510 15B

FPR × FNR

x = FPR, y = FNR · lower = better · ★ ideal (0,0)
0.000.000.250.250.500.500.750.751.001.00FPR →FNR →
★ Ideal (0,0) - zero errors

Heatmap fnr

min 0.004 max 0.721
FNR · group / model
YuFeng-XGuard-Reason-8B 8B
Sing-Guard-2b 2B
OpenGuardrails-Text-2510 15B
HiveTraceGuard-Pro 0.6B
Shieldstral-1.0-3B 3B
AEGIS 2.0 (prompt)
AEGIS 2.0 (response)
ToxicChat
PolyGuard (EN, request)
PolyGuard (EN, response)
PolyGuard (RU, request)
PolyGuard (RU, response)
RTP-LX (EN, request)
RTP-LX (RU, request)
RTP-LX (EN, response)
RTP-LX (RU, response)
XSTest
BeaverTails (response)
HarmBench (responses)
XSafety (EN)
OpenAI Moderation
Robustness Test — real (combined)
Robustness Test — responses real (combined)
Robustness Test — robust (combined)

Grouped bars

Recall / Precision / F1 by model · averages across selected groups
0.000.250.500.751.00
YuFeng-XGuard-Reason-8B 8B
Sing-Guard-2b 2B
HiveTraceGuard-Pro 0.6B
Shieldstral-1.0-3B 3B
OpenGuardrails-Text-2510 15B
Recall (higher = better) Precision (higher = better) F1 (higher = better)

Pareto: quality × latency

x = p95 latency (ms, lower = better), y = integral (higher = better)
0204060801001200.700.800.901.00p95 ms ↓Integral ↑
hover for details · dashed line = Pareto frontier

Latency performance

p50 / p95 / p99 (ms) · error rate
HiveTraceGuard-Pro 0.6B14.328.837.60.0%
Shieldstral-1.0-3B 3B14.732.244.90.0%
Sing-Guard-2b 2B62.180.8102.50.0%
YuFeng-XGuard-Reason-8B 8B71.8102.6144.30.0%
OpenGuardrails-Text-2510 15B63.2127.1181.30.0%

Single-request latency percentiles.

Robustness real → robust

degradation under augmentation - who breaks on harder attacks

Delta = robust - real. Delta score < 0: the model is not robust to obfuscations, > 0: robust. Delta FNR > 0: it misses harmful obfuscated messages more often (worse), < 0: it catches them more often (better). Delta FPR > 0: it blocks safe obfuscated messages more often (worse), < 0: it lets them pass more often (better).

Modelscore realscore robustΔ scoreFNR realFNR robustΔ FNRFPR realFPR robustΔ FPR
HiveTraceGuard-Pro 0.6B
0.9680.870−0.0980.0460.128+0.0820.0170.132+0.115
Sing-Guard-2b 2B
0.9310.849−0.0820.1190.241+0.1220.0120.036+0.024
Shieldstral-1.0-3B 3B
0.9090.834−0.0750.1330.238+0.1050.0440.078+0.035
YuFeng-XGuard-Reason-8B 8B
0.9330.870−0.0630.1040.214+0.1090.0260.027+0.001
OpenGuardrails-Text-2510 15B
0.8040.741−0.0630.3250.399+0.0740.0060.033+0.027