tim-lawson commited on
Commit
113a77e
1 Parent(s): 5b5d459

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +38 -15
README.md CHANGED
@@ -1,25 +1,48 @@
1
  ---
 
2
  language: en
3
  library_name: mlsae
4
  license: mit
5
  tags:
6
- - model_hub_mixin
7
- - pytorch_model_hub_mixin
8
- datasets:
9
- - monology/pile-uncopyrighted
10
  ---
11
 
12
- # mlsae-pythia-160m-deduped-x8-k32-tfm
13
 
14
- A Multi-Layer Sparse Autoencoder (MLSAE) trained on the residual stream
15
- activation vectors from every layer of
16
- [EleutherAI/pythia-160m-deduped](https://huggingface.co/EleutherAI/pythia-160m-deduped)
17
- with an expansion factor of 8 and k = 32, over 1 billion tokens from
18
- [monology/pile-uncopyrighted](https://huggingface.co/datasets/monology/pile-uncopyrighted).
19
- This model includes the underlying transformer.
20
 
21
- For more details, see:
22
 
23
- - Paper: <https://arxiv.org/abs/2409.04185>
24
- - GitHub repository: <https://github.com/tim-lawson/mlsae>
25
- - Weights & Biases project: <https://wandb.ai/timlawson-/mlsae>
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ base_model: EleutherAI/pythia-160m-deduped
3
  language: en
4
  library_name: mlsae
5
  license: mit
6
  tags:
7
+ - arxiv:2409.04185
8
+ - model_hub_mixin
9
+ - pytorch_model_hub_mixin
 
10
  ---
11
 
12
+ # Model Card for tim-lawson/mlsae-pythia-160m-deduped-x8-k32-tfm
13
 
14
+ A Multi-Layer Sparse Autoencoder (MLSAE) trained on the residual stream activation
15
+ vectors from [EleutherAI/pythia-160m-deduped](https://huggingface.co/EleutherAI/pythia-160m-deduped) with an
16
+ expansion factor of R = 8 and sparsity k = 32, over 1 billion
17
+ tokens from [monology/pile-uncopyrighted](https://huggingface.co/datasets/monology/pile-uncopyrighted).
 
 
18
 
 
19
 
20
+ This model is a PyTorch Lightning MLSAETransformer module, which includes the underlying
21
+ transformer.
22
+
23
+
24
+ ### Model Sources
25
+
26
+ - **Repository:** <https://github.com/tim-lawson/mlsae>
27
+ - **Paper:** <https://arxiv.org/abs/2409.04185>
28
+ - **Weights & Biases:** <https://wandb.ai/timlawson-/mlsae>
29
+
30
+ ## Citation
31
+
32
+ **BibTeX:**
33
+
34
+ ```bibtex
35
+ @misc{lawson_residual_2024,
36
+ title = {Residual {{Stream Analysis}} with {{Multi-Layer SAEs}}},
37
+ author = {Lawson, Tim and Farnik, Lucy and Houghton, Conor and Aitchison, Laurence},
38
+ year = {2024},
39
+ month = oct,
40
+ number = {arXiv:2409.04185},
41
+ eprint = {2409.04185},
42
+ primaryclass = {cs},
43
+ publisher = {arXiv},
44
+ doi = {10.48550/arXiv.2409.04185},
45
+ urldate = {2024-10-08},
46
+ archiveprefix = {arXiv}
47
+ }
48
+ ```