noanabeshima's picture
Update README.md
b77c8ca verified
|
raw
history blame
763 Bytes
metadata
license: mit

Cached SAE/transcoder acts stored in CSR format. Not especially optimized for others' use/fleshed out.

If you want to use them, do

import torch

def load_feat_acts(fname):
  csr_kwargs = torch.load(fname)

  # The matrices are stored in space-efficient formats that're incompatible with torch's sparse csr tensor.
  # Convert them back before constructing the matrix.
  csr_kwargs['crow_indices'] = csr_kwargs['crow_indices'].int()
  csr_kwargs['col_indices'] = csr_kwargs['crow_indices'].int()
  csr_kwargs['values'] = csr_kwargs['values'].float()/255
  
  feat_acts = torch.sparse_csr_tensor(**csr_kwargs)
  return feat_acts

The activations are for the train split in https://huggingface.co/datasets/noanabeshima/TinyModelTokIds