Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,41 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Ethics Probes: pythia-7b
|
| 2 |
+
|
| 3 |
+
This repository contains Sparse Autoencoder (SAE) + Probe models trained on various ethics datasets.
|
| 4 |
+
|
| 5 |
+
## Model
|
| 6 |
+
- **Base Model**: EleutherAI/pythia-6.9b
|
| 7 |
+
- **Short Name**: pythia-7b
|
| 8 |
+
|
| 9 |
+
## Datasets
|
| 10 |
+
|
| 11 |
+
The following datasets have been processed:
|
| 12 |
+
|
| 13 |
+
- **[utilitarianism](./utilitarianism/)**: Best Layer: 18, Val Acc: 0.6719, Test Acc: 0.6623
|
| 14 |
+
|
| 15 |
+
## Repository Structure
|
| 16 |
+
|
| 17 |
+
```
|
| 18 |
+
.
|
| 19 |
+
├── utilitarianism/
|
| 20 |
+
│ ├── layer_0/
|
| 21 |
+
│ │ ├── sae.pt
|
| 22 |
+
│ │ ├── probe.pt
|
| 23 |
+
│ │ ├── config.json
|
| 24 |
+
│ │ └── metrics.json
|
| 25 |
+
│ ├── layer_1/
|
| 26 |
+
│ ├── ...
|
| 27 |
+
│ ├── best/ # Best performing layer
|
| 28 |
+
│ └── README.md # Dataset-specific details
|
| 29 |
+
└── README.md # This file
|
| 30 |
+
```
|
| 31 |
+
|
| 32 |
+
## Usage
|
| 33 |
+
|
| 34 |
+
Each dataset folder contains:
|
| 35 |
+
- `layer_N/`: Models for each layer
|
| 36 |
+
- `best/`: Copy of the best performing layer's models
|
| 37 |
+
- `sae.pt`: Sparse Autoencoder state dict
|
| 38 |
+
- `probe.pt`: Probe state dict
|
| 39 |
+
- `config.json`: Configuration used for training
|
| 40 |
+
- `metrics.json`: Performance metrics
|
| 41 |
+
- `README.md`: Detailed results for the dataset
|