launch/MET-D-Gemma3-4B
MET-D-Gemma3-4B is a multilingual moral reasoning model fine-tuned from Google's Gemma-3-4B-it. This 4 billion parameter model is designed to judge actions from a specific character's perspective within a moral dilemma, providing a chain-of-thought explanation. It supports six languages (English, Spanish, Hindi, Korean, Malay, Chinese) and generates both reasoning and answers in the prompt's language. The model is optimized for nuanced ethical evaluations and cross-cultural moral reasoning.
Loading preview...
Overview
MET-D-Gemma3-4B is a specialized multilingual moral reasoning model, fine-tuned from the google/gemma-3-4b-it base model. Its core function is to evaluate a candidate action within a given moral dilemma, from the perspective of a described character. The model provides a judgment (Yes/No/Ambiguous) on the action's acceptability and assesses potential emotional discomfort for the character, complete with an explicit chain-of-thought explanation.
Key Capabilities
- Multilingual Moral Reasoning: Trained on self-generated reasoning traces across six languages: English, Spanish, Hindi, Korean, Malay, and Chinese. It generates reasoning and answers in the language of the prompt.
- Character-Perspective Judgment: Judges actions based on a provided character description, offering a nuanced ethical evaluation.
- Chain-of-Thought Explanations: Provides detailed reasoning for its judgments, making the decision-making process transparent.
- Rejection-Sampling for Accuracy: Utilizes rejection-sampling against ground-truth answers, conditioned on theoretical grounds specific to language and situation, to refine its reasoning traces.
Use Cases
This model is particularly suited for applications requiring:
- Automated ethical analysis from diverse viewpoints.
- Cross-cultural moral decision-making simulations.
- Educational tools for exploring ethical dilemmas.
- Content moderation requiring context-aware moral judgments.
Model Variants
MET-D-Gemma3-4B is part of a larger MET collection that includes variants based on different base models (e.g., Qwen3-4B, Qwen3-8B) and single-language specific checkpoints, offering flexibility for various deployment needs.