Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,54 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
language:
|
| 4 |
+
- en
|
| 5 |
+
tags:
|
| 6 |
+
- mechanistic interpretability
|
| 7 |
+
- sparse autoencoder
|
| 8 |
+
- llama
|
| 9 |
+
- llama-3
|
| 10 |
+
---
|
| 11 |
+
|
| 12 |
+
## Model Information
|
| 13 |
+
|
| 14 |
+
A SAE (Sparse Autoencoder) for [deepseek-ai/DeepSeek-R1-Distill-Llama-8B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B).
|
| 15 |
+
|
| 16 |
+
It is trained specifically on layer 19 of DeepSeek-R1-Distill-Llama-8B and achieves a final L0 of 93 during training.
|
| 17 |
+
|
| 18 |
+
This model is used to decompose Llama's activations into interpretable features.
|
| 19 |
+
|
| 20 |
+
The SAE weights are released under Apache, however DeepSeek-R1-Distill-Llama-8B is to be used under Meta's Llama 3 License.
|
| 21 |
+
|
| 22 |
+
## How to use
|
| 23 |
+
|
| 24 |
+
A Jupyter Notebook is provided to test the model
|
| 25 |
+
|
| 26 |
+
<a target="_blank" href="https://colab.research.google.com/github/qrsch/SAE/blob/main/SAE.ipynb">
|
| 27 |
+
<img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab", width="200px"/>
|
| 28 |
+
</a>
|
| 29 |
+
|
| 30 |
+
## Training
|
| 31 |
+
|
| 32 |
+
Our SAE was trained using [LMSYS-Chat-1M dataset](https://arxiv.org/pdf/2309.11998).
|
| 33 |
+
|
| 34 |
+
## Acknowledgements
|
| 35 |
+
|
| 36 |
+
This release wouldn't have been possible without the work of [Goodfire](https://www.goodfire.ai/) and [Anthropic](https://transformer-circuits.pub/)
|
| 37 |
+
|
| 38 |
+
```
|
| 39 |
+
.x+=:.
|
| 40 |
+
z` ^% .uef^"
|
| 41 |
+
.u . . <k .u . :d88E
|
| 42 |
+
.u@u .d88B :@8c .u .@8Ned8" .u u .d88B :@8c . `888E
|
| 43 |
+
.zWF8888bx ="8888f8888r ud8888. .@^%8888" ud8888. us888u. ="8888f8888r .udR88N 888E .z8k
|
| 44 |
+
.888 9888 4888>'88" :888'8888. x88: `)8b. :888'8888. .@88 "8888" 4888>'88" <888'888k 888E~?888L
|
| 45 |
+
I888 9888 4888> ' d888 '88%" 8888N=*8888 d888 '88%" 9888 9888 4888> ' 9888 'Y" 888E 888E
|
| 46 |
+
I888 9888 4888> 8888.+" %8" R88 8888.+" 9888 9888 4888> 9888 888E 888E
|
| 47 |
+
I888 9888 .d888L .+ 8888L @8Wou 9% 8888L 9888 9888 .d888L .+ 9888 888E 888E
|
| 48 |
+
`888Nx?888 ^"8888*" '8888c. .+ .888888P` '8888c. .+ 9888 9888 ^"8888*" ?8888u../ 888E 888E
|
| 49 |
+
"88" '888 "Y" "88888% ` ^"F "88888% "888*""888" "Y" "8888P' m888N= 888>
|
| 50 |
+
88E "YP' "YP' ^Y" ^Y' "P' `Y" 888
|
| 51 |
+
98> J88"
|
| 52 |
+
'8 @%
|
| 53 |
+
` :"
|
| 54 |
+
```
|