ss_bridges_d3072_f0.0039

ss_d3072_f0.0039

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

This repo contains a sparse model and bridges that couple it to a frozen dense model.

Sparse Model Details

• Layers: 2
• Model Dimension: 3072
• Context Length: 512
• Head Dimension: 16
• Vocabulary Size: None

Dense Model

• Source: jacobcd52/ss_d128_f1

Bridges

• Encoder AbsTopK Fraction: 0.25

Sparsity

• Weight Sparsity: True
• Target L0 Fraction: 0.00390625
• Activation Sparsity: True

Training

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

Usage

import torch
from huggingface_hub import hf_hub_download

# Download sparse model and bridges
sparse_model_path = hf_hub_download(repo_id="jacobcd52/ss_d3072_f0.0039", filename="sparse_model.bin")
bridges_path = hf_hub_download(repo_id="jacobcd52/ss_d3072_f0.0039", filename="bridges.bin")
config_path = hf_hub_download(repo_id="jacobcd52/ss_d3072_f0.0039", 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)