You can deploy this model and start earning money today!

markhenry/
msae-131k-3L-mlp_in-v1

Name: msae-131k-3L-mlp_in-v1
Availability: InStock

Views

No views yet

msae-131k-3L-mlp_in-v1

3-level Matryoshka BatchTopK SAE (dict=131072, k=48, groups=[1024,31744,98304]) trained on layer 6 mlp_in of vanilla-loss-parity-v2. Note: outer group had ~14% dead features due to high overcomplete ratio.

Training details

trainer: MatryoshkaBatchTopKTrainer
dict_size: 131072
k: 112
group_sizes: [1024, 31744, 98304]
activation_dim: 1024
layer: 6
location: mlp_in
base_model: hf_cache/vanilla-loss-parity-v2
steps: 97656
lr: 5e-05

Usage

python

from dictionary_learning.trainers.matryoshka_batch_top_k import MatryoshkaBatchTopKSAE
sae = MatryoshkaBatchTopKSAE.from_pretrained('ae.pt', device='cuda')
# encode: sae(x) returns (recon, f, loss_dict)
# decode: sae.decode(f)

Part of a research project comparing Matryoshka SAEs and CayleySAEs on hierarchy properties. See evals/msae/ in sparse-nanogpt-private.

markhenry/msae-131k-3L-mlp_in-v1

msae-131k-3L-mlp_in-v1

Training details

Usage

markhenry/
msae-131k-3L-mlp_in-v1