launch/MET-D-Gemma3-4B-en-only
MET-D-Gemma3-4B-en-only is an English-only moral reasoning model fine-tuned from Google's Gemma-3-4B-it. This 4 billion parameter model is designed to judge the acceptability of actions from a specific character's perspective within moral dilemmas, providing explicit chain-of-thought explanations. It specializes in generating reasoning traces and final judgments in English, based on self-generated, rejection-sampled training data. The model is optimized for nuanced ethical evaluations and understanding character-specific moral frameworks.
Loading preview...
Overview
MET-D-Gemma3-4B-en-only is a specialized English-only moral reasoning model, fine-tuned from Google's Gemma-3-4B-it. It is designed to analyze moral dilemmas by taking a specific character's perspective to judge the acceptability of an action and assess potential emotional discomfort. The model generates an explicit chain-of-thought explanation before providing a 'Yes', 'No', or 'Ambiguous' answer for action acceptability, and a 'Yes' or 'No' for discomfort.
Key Capabilities
- Character-Perspective Moral Judgment: Evaluates actions from a described character's viewpoint within a given moral dilemma.
- Chain-of-Thought Reasoning: Provides detailed explanations for its judgments, making the decision-making process transparent.
- Emotional Discomfort Assessment: Determines if performing or not performing an action would cause mental or emotional discomfort for the character.
- English-Only Focus: This specific variant is trained and optimized exclusively for English language moral reasoning tasks.
Training and Uniqueness
The model was trained using self-generated reasoning traces, which were then rejection-sampled against ground-truth answers for each character perspective. This method ensures that the model's reasoning aligns with established theoretical grounds. It is part of the broader MET collection which includes multilingual and other single-language variants, but this checkpoint is specifically tailored for English-language applications requiring nuanced ethical analysis.