A re-implementation of the Red Teaming Language Models with Language Models 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 -
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.
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.
The questions, answers and toxicity scores for each model can be found in the csv files in the artifacts directory.
For inference, I have used 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.