launch/MET-D-Qwen3-4B
MET-D-Qwen3-4B is a 4 billion parameter multilingual moral reasoning model fine-tuned from Qwen/Qwen3-4B. Developed by launch, it is designed to judge actions from a specific character's perspective in moral dilemmas, providing explicit chain-of-thought reasoning. This model excels at generating both reasoning traces and final judgments in six languages (English, Spanish, Hindi, Korean, Malay, Chinese), making it suitable for applications requiring nuanced, multilingual ethical AI. Its training involves self-generated reasoning traces rejection-sampled against ground truth for character perspectives.
Loading preview...
Overview
MET-D-Qwen3-4B is a 4 billion parameter multilingual moral reasoning model, fine-tuned from the Qwen/Qwen3-4B base model. Its primary function is to evaluate actions within moral dilemmas from a specified character's perspective, providing a judgment and a detailed chain-of-thought explanation. This model is part of the larger MET collection which explores moral reasoning across various base models and language subsets.
Key Capabilities
- Multilingual Moral Reasoning: Judges actions and provides reasoning in six languages: English, Spanish, Hindi, Korean, Malay, and Chinese.
- Perspective-Based Judgment: Answers two core questions from a character's viewpoint: whether an action is acceptable (Yes/No/Ambiguous) and whether doing/not doing it would cause discomfort (Yes/No).
- Explicit Chain-of-Thought: Generates detailed reasoning traces before providing a final answer, enhancing transparency and interpretability.
- Rejection-Sampling Training: Utilizes self-generated reasoning traces, rejection-sampled against ground truth based on character perspectives and theoretical grounds, to improve accuracy and alignment.
Use Cases
- Ethical AI Development: Ideal for research and applications requiring AI systems to navigate complex moral scenarios and provide reasoned judgments.
- Multilingual Content Analysis: Can be used to analyze ethical implications in text across diverse linguistic contexts.
- Educational Tools: Potentially useful in educational settings for exploring moral philosophy and decision-making.
Model Variants
This specific checkpoint combines all six languages. Single-language variants (e.g., launch/MET-D-Qwen3-4B-en-only) and variants based on other base models like Qwen3-8B and Gemma-3-4B are also available within the MET collection.