|
--- |
|
license: apache-2.0 |
|
language: |
|
- en |
|
tags: |
|
- mechanistic interpretability |
|
- sparse autoencoder |
|
- llama |
|
- llama-3 |
|
--- |
|
|
|
## Model Information |
|
|
|
A SAE (Sparse Autoencoder) for [meta-llama/Llama-3.2-1B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct). |
|
|
|
It is trained specifically on layer 9 of Llama 3.2 1B and achieves a final L0 of 63 during training. |
|
|
|
This model is used to decompose Llama's activations into interpretable features. |
|
|
|
The SAE weights are released under Apache, however Llama 3.2 1B is to be used under Meta's Llama 3.2 License. |
|
|
|
## How to use |
|
|
|
A Jupyter Notebook is provided to test the model |
|
|
|
<a target="_blank" href="https://colab.research.google.com/github/qrsch/SAE/blob/main/SAE.ipynb"> |
|
<img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab", width="200px"/> |
|
</a> |
|
|
|
## Training |
|
|
|
Our SAE was trained using [LMSYS-Chat-1M dataset](https://arxiv.org/pdf/2309.11998), on a single RTX 3090. The training script will be provided soon in the following repository: https://github.com/qrsch/SAE |
|
|
|
## Acknowledgements |
|
|
|
This release wouldn't have been possible without the work of [Goodfire](https://www.goodfire.ai/) and [Anthropic](https://transformer-circuits.pub/) |
|
|
|
``` |
|
.x+=:. |
|
z` ^% .uef^" |
|
.u . . <k .u . :d88E |
|
.u@u .d88B :@8c .u .@8Ned8" .u u .d88B :@8c . `888E |
|
.zWF8888bx ="8888f8888r ud8888. .@^%8888" ud8888. us888u. ="8888f8888r .udR88N 888E .z8k |
|
.888 9888 4888>'88" :888'8888. x88: `)8b. :888'8888. .@88 "8888" 4888>'88" <888'888k 888E~?888L |
|
I888 9888 4888> ' d888 '88%" 8888N=*8888 d888 '88%" 9888 9888 4888> ' 9888 'Y" 888E 888E |
|
I888 9888 4888> 8888.+" %8" R88 8888.+" 9888 9888 4888> 9888 888E 888E |
|
I888 9888 .d888L .+ 8888L @8Wou 9% 8888L 9888 9888 .d888L .+ 9888 888E 888E |
|
`888Nx?888 ^"8888*" '8888c. .+ .888888P` '8888c. .+ 9888 9888 ^"8888*" ?8888u../ 888E 888E |
|
"88" '888 "Y" "88888% ` ^"F "88888% "888*""888" "Y" "8888P' m888N= 888> |
|
88E "YP' "YP' ^Y" ^Y' "P' `Y" 888 |
|
98> J88" |
|
'8 @% |
|
` :" |
|
``` |
