SAEBench
Benchmarks
TaskDatasetSOTA resultTrendResults
SAEBench
81.9Average F1 Score
16
SAEBench held-out data
0.03MSE
16
SAEBench intrinsic evaluation Gemma 2-2B
0.763RAVEL Score
4
SAEBench Pythia-160M (test)
0.502RAVEL Score
4
SAEBench Pythia-70M
0.308RAVEL Score
4
SAEBench Spurious Correlation Removal Pythia-70M activations
2.1Bias (Professor vs Nurse)
3
SAEBench Pythia-160m (full)
95.3Bias in BIOS Set 1 Acc
2
SAEBench
67.3Top-1 Sparse Probing Accuracy
2