Skip to content

Latest commit

 

History

History
11 lines (8 loc) · 762 Bytes

README.md

File metadata and controls

11 lines (8 loc) · 762 Bytes
Raw

GPT-3.5 Saliency Map Verbalizations

alt text

Run

  1. In ChatGPT_class.py, change the api_key = "" and organization = "" (l. 7) to your OpenAI credentials.
  2. SMV_Lists.py includes the IDs for the IMDb and AG News datasets that we used in our human evaluation. You can change it to your desired data points, but I would advise not to put the entire data in there, because it will not be cheap. ; )
  3. Decide on your OpenAI model of choice in ChatGPT_test.py (l. 8). At the time when we conducted our study, "gpt-3.5-turbo" was the best one available.
  4. python ChatGPT_test.py will produce the rationales based on the IDs in SMV_Lists.py and write them to Weights & Biases (you may need to log in first).