ss_bridges_d1024_f0.015625_resid

ss_bridges_d1024_f0.015625_resid

Weight-sparse transformer with bridges, trained with the procedure from Gao et al. (2025).

Model Details (Sparse Model)

• Layers: 2
• Model Dimension: 1024
• Context Length: 512
• Head Dimension: 16
• Vocabulary Size: 4096

Bridges

• Dense Model: jacobcd52/ss_d128_f1
• Encoder Activation Fraction: 0.25

Sparsity

• Weight Sparsity: True
• Target L0 Fraction: 0.015625
• Activation Sparsity: True

Training

• Dataset: data/simplestories-tokenized
• Tokenizer: SimpleStories/SimpleStories-1.25M
• Total Tokens: 2,000,000,000

Training Run

• W&B Run: https://wandb.ai/training-saes/bridges_training/runs/0g7bw4i5

Usage

import torch
from huggingface_hub import hf_hub_download

# Download model and bridges
sparse_model_path = hf_hub_download(repo_id="jacobcd52/ss_bridges_d1024_f0.015625_resid", filename="sparse_model.bin")
bridges_path = hf_hub_download(repo_id="jacobcd52/ss_bridges_d1024_f0.015625_resid", filename="bridges.bin")
config_path = hf_hub_download(repo_id="jacobcd52/ss_bridges_d1024_f0.015625_resid", filename="config.json")

# Load (requires the SparseGPT and BridgeSet classes from this repo)
sparse_state_dict = torch.load(sparse_model_path)
bridges_state_dict = torch.load(bridges_path)