tim-lawson
commited on
Commit
•
f707dbf
1
Parent(s):
07c2361
Push model using huggingface_hub.
Browse files- README.md +47 -0
- config.json +20 -0
- model.safetensors +3 -0
README.md
ADDED
@@ -0,0 +1,47 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
language: en
|
3 |
+
library_name: mlsae
|
4 |
+
license: mit
|
5 |
+
tags:
|
6 |
+
- arxiv:2409.04185
|
7 |
+
- model_hub_mixin
|
8 |
+
- pytorch_model_hub_mixin
|
9 |
+
---
|
10 |
+
|
11 |
+
# Model Card for tim-lawson/sae-pythia-410m-deduped-x64-k32-tfm-layers-21
|
12 |
+
|
13 |
+
A Multi-Layer Sparse Autoencoder (MLSAE) trained on the residual stream activation
|
14 |
+
vectors from [EleutherAI/pythia-410m-deduped](https://huggingface.co/EleutherAI/pythia-410m-deduped) with an
|
15 |
+
expansion factor of R = 64 and sparsity k = 32, over 1 billion
|
16 |
+
tokens from [monology/pile-uncopyrighted](https://huggingface.co/datasets/monology/pile-uncopyrighted).
|
17 |
+
|
18 |
+
|
19 |
+
This model is a PyTorch Lightning MLSAETransformer module, which includes the underlying
|
20 |
+
transformer.
|
21 |
+
|
22 |
+
|
23 |
+
### Model Sources
|
24 |
+
|
25 |
+
- **Repository:** <https://github.com/tim-lawson/mlsae>
|
26 |
+
- **Paper:** <https://arxiv.org/abs/2409.04185>
|
27 |
+
- **Weights & Biases:** <https://wandb.ai/timlawson-/mlsae>
|
28 |
+
|
29 |
+
## Citation
|
30 |
+
|
31 |
+
**BibTeX:**
|
32 |
+
|
33 |
+
```bibtex
|
34 |
+
@misc{lawson_residual_2024,
|
35 |
+
title = {Residual {{Stream Analysis}} with {{Multi-Layer SAEs}}},
|
36 |
+
author = {Lawson, Tim and Farnik, Lucy and Houghton, Conor and Aitchison, Laurence},
|
37 |
+
year = {2024},
|
38 |
+
month = oct,
|
39 |
+
number = {arXiv:2409.04185},
|
40 |
+
eprint = {2409.04185},
|
41 |
+
primaryclass = {cs},
|
42 |
+
publisher = {arXiv},
|
43 |
+
doi = {10.48550/arXiv.2409.04185},
|
44 |
+
urldate = {2024-10-08},
|
45 |
+
archiveprefix = {arXiv}
|
46 |
+
}
|
47 |
+
```
|
config.json
ADDED
@@ -0,0 +1,20 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"accumulate_grad_batches": 64,
|
3 |
+
"auxk": 256,
|
4 |
+
"auxk_coef": 0.03125,
|
5 |
+
"batch_size": 1,
|
6 |
+
"dead_steps_threshold": null,
|
7 |
+
"dead_threshold": 0.001,
|
8 |
+
"dead_tokens_threshold": 10000000,
|
9 |
+
"expansion_factor": 64,
|
10 |
+
"k": 32,
|
11 |
+
"layers": [
|
12 |
+
21
|
13 |
+
],
|
14 |
+
"lr": 0.0001,
|
15 |
+
"max_length": 2048,
|
16 |
+
"model_name": "EleutherAI/pythia-410m-deduped",
|
17 |
+
"skip_special_tokens": true,
|
18 |
+
"standardize": true,
|
19 |
+
"tuned_lens": false
|
20 |
+
}
|
model.safetensors
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:a7b1bf8515b07ebaf27ada1df39a8510adbd0525700dc812b1122583352794b3
|
3 |
+
size 2158251016
|