Tomasdulka's workspace
Runs
3
State
Notes
User
Tags
Created
Runtime
Sweep
activation_dim
dataset
device
dict_class
dict_size
dictionary_size
k
layer
lm_name
lr
model
seed
steps
trainer_class
training_approach
training_approaches
wandb_name
auxk_loss
dead_features
effective_l0
frac_variance_explained
l0
l2_loss
loss
Finished
tomasdulka
12m 11s
-
512
pile-uncopyrighted
cuda
AutoEncoderTopK
32768
32768
100
3
EleutherAI/pythia-70m-deduped
0.00014142
EleutherAI/pythia-70m-deduped
0
30000
TrainerTopK
-
["topk","batch_topk_to_jump","jumprelu"]
topk_sae_run
0
0
100
0.93892
100
4.99742
4.99742
Finished
tomasdulka
14m 18s
-
512
pile-uncopyrighted
cuda
JumpReluAutoEncoder
32768
32768
-
3
EleutherAI/pythia-70m-deduped
0.00007
EleutherAI/pythia-70m-deduped
0
30000
TrainerJumpRelu
jumprelu
-
jumprelu_sae_run
-
-
-
0.94015
190.40625
5.06815
19.00643
Finished
tomasdulka
16m 30s
-
512
pile-uncopyrighted
cuda
BatchTopKToJumpSAE
32768
32768
100
3
EleutherAI/pythia-70m-deduped
0.00014142
EleutherAI/pythia-70m-deduped
0
30000
TrainerBatchTopK
batch_topk_to_jump
-
batch_topk_sae_run
0.00027919
-1
100
0.94182
100
4.57427
4.57428
1-3
of 3