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

Gemma-2

Benchmarks

Task NameDataset NameSOTA ResultTrend
Watermark Robustness AnalysisGemma-2 2B
Post-attack TPR100
49
Semantic similarity analysisGemma-2 9B within-prompt completions
Cosine Distance0.312
8
Semantic similarity analysisGemma-2 within-prompt completions 2B
Cosine Distance0.316
8
Input-based feature description evaluationGemma-2 MLP SAE features
Score56.6
8
Input-based feature description evaluationGemma-2 Residual SAE features
Feature Description Score67
8
Watermark Detection RobustnessGemma-2 9B Pre-trained (PT) (test)
TPR (Baseline)100
7
Watermarked text generation and detectionGemma-2 2B-IT
TPR99
1
Showing 7 of 7 rows