Dataset: ETHICS at Hugging Face
This dataset holds informations on whether a scenario is ethical or not. It includes two columns:
| label | scenario |
|---|---|
| 1 or 0 | text |
... where 1 means the scenario is ethical, and 0 means the scenario is not ethical.
There are 5 subsets available:
| subset | rows |
|---|---|
| Justice | 26.5K |
| Common Sense | 21.8K |
| Utilitarianism | 22.8K |
| Deontology | 25.3K |
| Virtue | 38K |
Released on: 2020.
Link: Full Paper
This foundational paper introduces the ETHICS dataset, a benchmark designed to evaluate a language model's understanding of five core domains of normative ethics. By testing various baseline neural networks, the authors establish the initial performance metrics and demonstrate that modern models still struggle significantly with nuanced moral reasoning.
Released on: 2025.
Link: Full Paper
This study develops a specialized framework called Ethic-BERT, utilizing advanced optimization techniques like full model unfreezing and adaptive learning rates to improve ethical text classification. The authors benchmark their model against the original dataset's challenging "Hard Test" split, successfully outperforming previous baselines and providing a stronger model for comparison.
Released on: 2023.
Link: Full Paper
This paper expands the original benchmark by adapting the ETHICS dataset into a multi-label framework where scenarios are evaluated across all five ethical principles simultaneously. The authors introduce the Ethical Alignment Language Model (EALM) to prove that neural networks can be trained to handle real-world, multi-perspective moral reasoning rather than just isolated ethical tasks.
-
Review each article in depth and summarize the experiments and highlight struggles.
-
Redo the experiments and get the same results.
-
Trying on new experiments to get better models.