Goal-conditioned Latent Factor Discovery without Task Supervision

---

Steps to run:

• Navigate to each task's folder
• Run generate_concepts.py to propose attributes with LLM (need openai API key, or use open-source models)
  • See e.g., this code block for how to switch models and adjust multi-processing num threads
  • The concept annotation process is parallelized thanks to Dongfu Jiang's framework
• Run train_nli.py to train compatibility estimation model
• Run fit_corex.py to learn latent factor model
• Run evaluate.py to get evaluation results (dataset/use-case specific, really)
• Questions about the environment? See requirements.md

Using your own dataset?

• The main part to modify is generate_concepts.py, specifically:
  • new data loading scripts
  • modify the instructions to LLMs to fit your task