These are interesting but ultimately throwaway numbers. However, I think benchmarks like these suggest that the value provided by giving your data to openai can outweigh the costs. Prompt engineering provides an avenue to prevent leaking sensitive data into the system while still creating value with openai projects
| llm model | false alarms | detections |
|---|---|---|
| LLaMa 65B (4-bit GPTQ) | 1 / 15 | 0 / 13 |
| Baize 30B (8-bit) | 0 / 15 | 1 / 13 |
| Galpaca 30B (8-bit) | 0 / 15 | 1 / 13 |
| Koala 13B (8-bit) | 0 / 15 | 0 / 13 |
| Vicuna 13B (8-bit) | 2 / 15 | 1 / 13 |
| Vicuna 7B (FP16) | 1 / 15 | 0 / 13 |
| GPT 3.5 | 0 / 15 | 7 / 13 |
| GPT 4 | 0 / 15 | 13 / 13 |