https://github.com/seonglae/emgsd-hermes

Steering GPT2-EMGSD less biased & Generating stereotyped text with vanilla GPT2 without fine tuning or prompt engineering
https://github.com/seonglae/emgsd-hermes

bias-correction bias-mitigation emgsd gpt2 sparse-autoencoder steering-vector stereotype

Last synced: 24 days ago
JSON representation

Steering GPT2-EMGSD less biased & Generating stereotyped text with vanilla GPT2 without fine tuning or prompt engineering

• Host: GitHub
• URL: https://github.com/seonglae/emgsd-hermes
• Owner: seonglae
• Created: 2024-11-23T10:10:36.000Z (about 1 month ago)
• Default Branch: main
• Last Pushed: 2024-12-01T02:35:20.000Z (about 1 month ago)
• Last Synced: 2024-12-01T03:24:34.497Z (about 1 month ago)
• Topics: bias-correction, bias-mitigation, emgsd, gpt2, sparse-autoencoder, steering-vector, stereotype
• Language: Jupyter Notebook
• Homepage: https://www.canva.com/design/DAGXTdRh__E/MfKS_-4Px4iXNyCT89Sh5g/edit
• Size: 425 KB
• Stars: 2
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# EMGSD Hermes

This project explores bias mitigation in GPT2-EMGSD, leveraging correlation analysis for stereotype deduction and activation manipulation, highlighting the potential of an alternative to traditional fine-tuning. Additionally, it demonstrates the feasibility of inducing bias in vanilla GPT2 through activation engineering.



## Fast Demo

```bash

# Install python 3.10 which is required by SAE-Lens

⁠⁠⁠git clone ⁠ https://github.com/seonglae/emgsd-hermes && cd emgsd-hermes

p⁠ip install torch colorama sae-lens transformers

python compare.py

```

## Main Pipeline

TBA

### 1. Fine-tuning SAE with EMGSD dataset 

```bash

python empsd.py

```

### 2. Extract features using correlation

```bash

python search_category.py

python search_stereo.py

# replace emgsd/*.json files

python draw_corr.py

```



or if you want to calculate mutual information

```

python mi_stereo.py

```

### 3. Compute ratio of stereotyped text in generation

```bash

python compare_all.py

```





## Loss Graph of fine-tuning SAE