leukas/hippo_superset

The HIPPO 🦛 Superset

This repository contains our training and evaluation superset for HIPPO: Harmful Input, Positive and Productive Output. This is part of our work, From Specialization to Generalization: Instruction-tuned LLMs for Robust Harmful Content Mitigation, where we show that instruction-tuning an LLM for hate speech can dramatically improve its performance on hate speech-related tasks. We've collected 36 different hate speech datasets and converted them into a conversational template. We used it to train our HIPPO 4B and 32B models, which outperform several task-specific models.

• Our code: https://github.com/TUM-NLP/HIPPO
• Our 4B model: leukas/hippo-4b
• Our 32B model: leukas/hippo-32b