Residual Stream Analysis with Multi-Layer SAEs
Paper
•
2409.04185
•
Published
•
1
Single-Layer SAEs
TopK SAEs trained on the residual stream activation vectors from a single transformer layer.
-
-
tim-lawson/sae-pythia-70m-deduped-x64-k32-layers-0
0.0B • Updated • 5 -
tim-lawson/sae-pythia-70m-deduped-x64-k32-layers-1
0.0B • Updated • 3 -
tim-lawson/sae-pythia-70m-deduped-x64-k32-layers-2
0.0B • Updated • 3 -
tim-lawson/sae-pythia-70m-deduped-x64-k32-layers-3
0.0B • Updated • 3 -
tim-lawson/sae-pythia-70m-deduped-x64-k32-layers-4
0.0B • Updated • 5 -
tim-lawson/sae-pythia-70m-deduped-x64-k32-layers-5
0.0B • Updated • 5 -
tim-lawson/sae-pythia-160m-deduped-x64-k32-layers-0
0.1B • Updated • 3 -
tim-lawson/sae-pythia-160m-deduped-x64-k32-layers-1
0.1B • Updated • 3 -
tim-lawson/sae-pythia-160m-deduped-x64-k32-layers-2
0.1B • Updated • 3 -
tim-lawson/sae-pythia-160m-deduped-x64-k32-layers-3
0.1B • Updated • 5 -
tim-lawson/sae-pythia-160m-deduped-x64-k32-layers-4
0.1B • Updated • 3 -
tim-lawson/sae-pythia-160m-deduped-x64-k32-layers-5
0.1B • Updated • 3 -
tim-lawson/sae-pythia-160m-deduped-x64-k32-layers-6
0.1B • Updated • 3 -
tim-lawson/sae-pythia-160m-deduped-x64-k32-layers-7
0.1B • Updated • 3 -
tim-lawson/sae-pythia-160m-deduped-x64-k32-layers-8
0.1B • Updated • 3 -
tim-lawson/sae-pythia-160m-deduped-x64-k32-layers-9
0.1B • Updated • 3 -
tim-lawson/sae-pythia-160m-deduped-x64-k32-layers-10
0.1B • Updated • 3 -
tim-lawson/sae-pythia-160m-deduped-x64-k32-layers-11
0.1B • Updated • 7 -
tim-lawson/sae-pythia-410m-deduped-x64-k32-layers-0
0.1B • Updated • 3 -
tim-lawson/sae-pythia-410m-deduped-x64-k32-layers-1
0.1B • Updated • 4 -
tim-lawson/sae-pythia-410m-deduped-x64-k32-layers-2
0.1B • Updated • 3 -
tim-lawson/sae-pythia-410m-deduped-x64-k32-layers-3
0.1B • Updated • 5 -
tim-lawson/sae-pythia-410m-deduped-x64-k32-layers-4
0.1B • Updated • 3 -
tim-lawson/sae-pythia-410m-deduped-x64-k32-layers-5
0.1B • Updated • 3 -
tim-lawson/sae-pythia-410m-deduped-x64-k32-layers-6
0.1B • Updated • 2 -
tim-lawson/sae-pythia-410m-deduped-x64-k32-layers-7
0.1B • Updated • 2 -
tim-lawson/sae-pythia-410m-deduped-x64-k32-layers-8
0.1B • Updated • 4 -
tim-lawson/sae-pythia-410m-deduped-x64-k32-layers-9
0.1B • Updated • 3 -
tim-lawson/sae-pythia-410m-deduped-x64-k32-layers-10
0.1B • Updated • 3 -
tim-lawson/sae-pythia-410m-deduped-x64-k32-layers-11
0.1B • Updated • 3 -
tim-lawson/sae-pythia-410m-deduped-x64-k32-layers-12
0.1B • Updated • 4 -
tim-lawson/sae-pythia-410m-deduped-x64-k32-layers-13
0.1B • Updated • 3 -
tim-lawson/sae-pythia-410m-deduped-x64-k32-layers-14
0.1B • Updated • 5 -
tim-lawson/sae-pythia-410m-deduped-x64-k32-layers-15
0.1B • Updated • 4 -
tim-lawson/sae-pythia-410m-deduped-x64-k32-layers-16
0.1B • Updated • 2 -
tim-lawson/sae-pythia-410m-deduped-x64-k32-layers-17
0.1B • Updated • 2 -
tim-lawson/sae-pythia-410m-deduped-x64-k32-layers-18
0.1B • Updated • 2 -
tim-lawson/sae-pythia-410m-deduped-x64-k32-layers-19
0.1B • Updated • 2 -
tim-lawson/sae-pythia-410m-deduped-x64-k32-layers-20
0.1B • Updated • 2 -
tim-lawson/sae-pythia-410m-deduped-x64-k32-layers-21
0.1B • Updated • 3 -
tim-lawson/sae-pythia-410m-deduped-x64-k32-layers-22
0.1B • Updated • 3 -
tim-lawson/sae-pythia-410m-deduped-x64-k32-layers-23
0.1B • Updated • 4