msae-gpt2-small-2L-v1

2-level Matryoshka BatchTopK SAE (dict=4096, k=32, groups=[768,3328]) trained on layer 6 mlp_in of gpt2 (small). Falsification run: tests whether inferred-hierarchy flatness is scale-specific.

Training details

• trainer: MatryoshkaBatchTopKTrainer
• dict_size: 4096
• k: 32
• group_sizes: [768, 3328]
• activation_dim: 768
• layer: 6
• location: mlp_in_layer_6
• base_model: gpt2
• steps: 14648
• lr: 5e-05

Usage

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.