README.md

metadata

license: apache-2.0

Goodtriever

This repository contains datasets and model generations from the Goodtriever: Adaptive Toxicity Mitigation with Retrieval-augmented Models paper, published as a conference paper on EMNLP 2023.

[Paper][Code][Data]

• data: prompts and datasets used for datastore creation.
  • continual_mitigation: clustered WILDS data and prompts
  • datastore_quality: for the experiments on how automatic labeling impacts mitigation results
  • jigsaw: main dataset, jigsaw unintended bias
• outputs: model generations and results for all experiments from the paper.
  • alpha_temperature
  • datastore_quality
  • datastore_size
  • k_neighbors
  • model_families (and main table results)

Citation