Add DecodingTrust Fairness dataset loader#2231
Conversation
Adds _DecodingTrustFairnessDataset, a remote loader for the Fairness perspective of the DecodingTrust benchmark (Adult census income classification, base-rate-parity variants). Scoped to Adult subsets only; COMPAS/crime data in the same source is a natural follow-up. Relates to microsoft#291
|
@microsoft-github-policy-service agree |
|
I've thought about this one for a while and there are lots of detailed problems here, e.g., whether this fits the "prompt" seed dataset or perhaps more a Q&A (FWIW I think it's neither... more of a eval); pinning the SHA; adding few-shot examples; rerunning 1_loading_datasets.ipynb, and others. I'd like to focus on the high-level setting, though. The Adult dataset is based on the census from 1994. That data is now 32 years old. Asking a model today whether someone is likely to earn more or less than $50k is kind of odd. Even further, I'm 99.999% sure every model out there has ingested this data anyway and so we're potentially just predicting whether it can reconstruct the dataset. If there was newer census data available I might think differently but I suspect this won't be useful. |
2 participants
Adds _DecodingTrustFairnessDataset, a remote loader for the Fairness perspective of the DecodingTrust benchmark (Adult census income classification, base-rate-parity variants). Scoped to Adult subsets only; COMPAS/crime data in the same source is a natural follow-up.
Relates to #291
Description
Tests and Documentation