FEAT: CBT-Bench Dataset

[sravan0446]

CBT-Bench Dataset Integration for PyRIT

This pull request introduces support for the CBT-Bench dataset in PyRIT. The CBT-Bench dataset is a benchmark designed to evaluate the alignment and therapeutic safety of Large Language Models (LLMs) in the context of Cognitive Behavioral Therapy (CBT). By integrating this dataset, PyRIT can analyze psychotherapy-related prompts and identify potential risks, including cognitive distortions, harmful self-beliefs, and other vulnerabilities.

Key Changes

• Added a new function fetch_cbt_bench_dataset to fetch and process the dataset.
• Mapped dataset fields (situation, core_belief_fine_grained) to PyRIT's SeedPrompt model.
• Included default configuration (core_fine_seed) with support for other configurations.
• Updated documentation and comments to explain the purpose and usage of the dataset.

Related Issue

Closes: #865
Tagging @romanlutz for review.

Tests and Documentation

• Tests: Unit tests have not been added. They can be included in a subsequent commit within the same pull request if requested by the reviewer.
• Documentation: Inline documentation has been updated. The README has not been updated yet but will be addressed if deemed necessary.
• JupyText: Not applicable, as this contribution focuses on dataset integration rather than API or documentation changes.

References

• CBT-Bench Dataset on Hugging Face

Citation

@article{zhang2024cbt,
  title={CBT-Bench: Evaluating Large Language Models on Assisting Cognitive Behavior Therapy},
  author={Zhang, Mian and Yang, Xianjun and Zhang, Xinlu and Labrum, Travis and Chiu, Jamie C and Eack, Shaun M and Fang, Fei and Wang, William Yang and Chen, Zhiyu Zoey},
  journal={arXiv preprint arXiv:2410.13218},
  year={2024}
}

[sravan0446]

@microsoft-github-policy-service agree

[romanlutz]

Please also look at api.rst (I think that's the name) because it should list this dataset along with the others from the datasets module. There may be some missing there, so if you want you can compare and fill in the gaps. Thank you!

[romanlutz]

This should tell us about the other options and the type should be Literal[...] with the individual options listed.

[romanlutz]

Would these overlap between the datasets (based on config value)? If so we need to distinguish

[romanlutz]

What are the values for this? I just want to make sure they make sense as harm categories

[bashirpartovi]

@romanlutz you can see them here FYI

But they look like this:

[ 
       "I am powerless, weak, vulnerable", 
       "I am needy", 
       "I am out of control" 
]

I am not certain if they exactly align with our harm categories

[romanlutz]

These ones certainly don't.

[romanlutz]

I would say these broadly fall under psycho-social harms. @jbolor21 was that what you thought when you added the item?

[jbolor21]

Yes! I was thinking broadly under psycho-social harms!

[jbolor21]

The labels they have are very specific to CBT and not quite aligned with our categories so just lumping them under psycho-social harms would probably make the most sense!

[romanlutz]

After looking at the dataset, I would say the situation + thoughts together should be the prompt. Let's ask @jbolor21 who suggested adding the dataset to chime in 🙂

[jbolor21]

Yes I think adding the thoughts is important to the situation since these two are what are extracted from the original text!

[romanlutz]

@sravan0446 are you still interested in addressing the issues? 🙂

[romanlutz]

closing in favor of #1411