|
--- |
|
license: mit |
|
--- |
|
|
|
OpenAI's GPT2-Small SAEs reformatted for easy loading from SAE Lens. |
|
|
|
Links |
|
- [Paper](https://cdn.openai.com/papers/sparse-autoencoders.pdf) |
|
- [Original File Loading](https://github.com/openai/sparse_autoencoder/blob/lg-training/sparse_autoencoder/paths.py) |
|
|
|
```python |
|
import torch |
|
from transformer_lens import HookedTransformer |
|
from sae_lens import SAE, ActivationsStore |
|
|
|
torch.set_grad_enabled(False) |
|
model = HookedTransformer.from_pretrained("gpt2-small") |
|
sae, cfg, sparsity = SAE.from_pretrained( |
|
"gpt2-small-resid-post-v5-32k", # to see the list of available releases, go to: https://github.com/jbloomAus/SAELens/blob/main/sae_lens/pretrained_saes.yaml |
|
"blocks.11.hook_resid_post" # change this to another specific SAE ID in the release if desired. |
|
) |
|
|
|
# For loading activations or tokens from the training dataset. |
|
activation_store = ActivationsStore.from_sae( |
|
model=model, |
|
sae=sae, |
|
streaming=True, |
|
# fairly conservative parameters here so can use same for larger |
|
# models without running out of memory. |
|
store_batch_size_prompts=8, |
|
train_batch_size_tokens=4096, |
|
n_batches_in_buffer=4, |
|
device=device, |
|
) |
|
|
|
``` |
