Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

IHEval

Benchmarks

Task NameDataset NameSOTA ResultTrend
Instruction followingIHEval
PLA86.7
21
Task ExecutionIHEval
Language Detection (Reference)100
12
Tool UseIHEval Overall Tool Use v1 (All)
Average Accuracy69.9
12
Slack UserIHEval v1 (Conflict)
Accuracy80
12
Slack UserIHEval v1 (Aligned)
Accuracy83
12
Slack UserIHEval v1 (Reference)
Accuracy94
12
Get WebpageIHEval v1 (Conflict)
Accuracy39.8
12
Get WebpageIHEval Aligned v1
Accuracy55.9
12
Get WebpageIHEval v1 (Reference)
Accuracy86
12
Rule FollowingIHEval Single-Turn
Accuracy (Reference)88.5
12
Rule FollowingIHEval Multi-Turn
Accuracy (Reference)89.8
12
Safety EvaluationIHEval Average 1.0
Average Accuracy66.9
12
Prompt HijackingIHEval Prompt Hijacking Conflict 1.0
Accuracy45
12
Prompt HijackingIHEval Prompt Hijacking - Alignment 1.0
Accuracy82.5
12
Prompt HijackingIHEval Prompt Hijacking 1.0 (Reference)
Accuracy97.5
12
Prompt ExtractionIHEval Prompt Extraction - Conflict 1.0
Accuracy59.6
12
Prompt ExtractionIHEval Prompt Extraction Alignment 1.0
Accuracy83.7
12
Prompt ExtractionIHEval Prompt Extraction 1.0 (Reference)
Accuracy96.9
12
Prompt Injection DetectionIHEval Tool-use
FPR0
6
Prompt Injection DetectionIHEval Rule-following
FPR0.01
6
Showing 20 of 20 rows