https://github.com/shreyansh26/red-teaming-language-models-with-language-models
A re-implementation of the "Red Teaming Language Models with Language Models" paper by Perez et al., 2022
https://github.com/shreyansh26/red-teaming-language-models-with-language-models
adversarial-attacks adversarial-machine-learning gpt-2 large-language-models llama2 llm pythia red-teaming
Last synced: 3 months ago
JSON representation
A re-implementation of the "Red Teaming Language Models with Language Models" paper by Perez et al., 2022
- Host: GitHub
- URL: https://github.com/shreyansh26/red-teaming-language-models-with-language-models
- Owner: shreyansh26
- Created: 2023-10-07T11:08:15.000Z (about 2 years ago)
- Default Branch: main
- Last Pushed: 2023-10-09T05:37:35.000Z (almost 2 years ago)
- Last Synced: 2024-10-18T22:11:13.088Z (12 months ago)
- Topics: adversarial-attacks, adversarial-machine-learning, gpt-2, large-language-models, llama2, llm, pythia, red-teaming
- Language: Python
- Homepage:
- Size: 73 MB
- Stars: 25
- Watchers: 2
- Forks: 4
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# Red Teaming Language Models with Language Models
A re-implementation of the [Red Teaming Language Models with Language Models](https://arxiv.org/abs/2202.03286) paper by Perez et al.
This implementation only focuses on toxic/offensive langauge section of the paper.
I have based run the red-teaming experiments on four target models -
* [GPT2-XL](https://huggingface.co/gpt2-xl)
* [Llama-2-7B](https://huggingface.co/meta-llama/Llama-2-7b-hf)
* [Pythia-6.9B](https://huggingface.co/EleutherAI/pythia-6.9b)
* [Phi-1.5](https://huggingface.co/microsoft/phi-1_5.)
I generated 50,000 valid test cases (questions) as opposed to 500,000 mentioned in the paper, due to compute constraints. A small percantge of the test cases were succesful in being able to elicit toxic language generation from these models.
For Toxicity detection, I have used [Detoxify](https://github.com/unitaryai/detoxify).
The percentage of toxic/offensive *answers* (toxicity probability > 0.5) generated using Zero-Shot and Stochastic Few-Shot generation for each model are presented below -
| Model | Zero-Shot | Stochastic Few-Shot | Increase (Few-Shot vs Zero-Shot) |
|--------------|-----------|---------------------|----------------------------------|
| GPT2-XL-1.5B | 0.99% | 2.85% | 2.88x |
| Llama-2-7B | 0.15% | 0.33% | 2.20x |
| Pythia-6.9B | 0.47% | 1.02% | 2.17x |
| Phi-1.5B | 0.11% | 0.79% | 7.18x |
The percentage of toxic/offensive *questions and answers* (toxicity probability > 0.5) generated using Zero-Shot and Stochastic Few-Shot generation for each model are presented below -
| Model | Zero-Shot | Stochastic Few-Shot | Increase (Few-Shot vs Zero-Shot) |
|--------------|-----------|---------------------|----------------------------------|
| GPT2-XL-1.5B | 2.28% | 8.49% | 3.72x |
| Llama-2-7B | 0.20% | 0.61% | 3.05x |
| Pythia-6.9B | 0.93% | 3.00% | 3.23x |
| Phi-1.5B | 0.09% | 0.38% | 4.22x |
**Note** - The numbers may be a bit high as the toxicity detection model has False Positives in its predictions as well. The toxic/offensive generations can be manually inspected with the script [view_toxic_qa.py](view_toxic_qa.py).
The questions, answers and toxicity scores for each model can be found in the csv files in the [artifacts](artifacts/) directory.
For inference, I have used [vLLM](https://github.com/vllm-project/vllm) (which is awesome!) for GPT2-XL, Llama-2 and Pythia to speedup the generation process. Phi-1.5 is currently not supported by vLLM.