launch/MET-D-Qwen3-4B-ko-only

TEXT GENERATIONConcurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 8, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

MET-D-Qwen3-4B-ko-only is a Korean-only moral reasoning model developed by launch, fine-tuned from Qwen3-4B. This model is specifically designed to judge actions from a character's perspective within moral dilemmas, providing a chain-of-thought explanation before delivering a judgment. It excels at generating both reasoning traces and final answers in Korean, making it ideal for applications requiring nuanced ethical analysis in a Korean linguistic context. The model addresses the challenge of verifying reasoning traces by using character perspectives to rejection-sample its own reasoning, conditioned on theoretical grounds.

Loading preview...

Overview

launch/MET-D-Qwen3-4B-ko-only is a specialized moral reasoning model, fine-tuned from the Qwen3-4B base model. Its core function is to analyze moral dilemmas from a specific character's viewpoint, evaluating a given action and explaining its judgment through a detailed chain-of-thought. A key innovation is its use of character perspectives to generate ground-truth answers, which are then used for rejection-sampling the model's own reasoning traces, ensuring higher verifiability in complex ethical scenarios.

Key Capabilities

  • Korean-only Moral Reasoning: Exclusively trained and optimized for processing and generating moral judgments and explanations in Korean.
  • Perspective-based Judgment: Assesses actions within moral dilemmas from a defined character's perspective, providing nuanced ethical analysis.
  • Chain-of-Thought Explanations: Generates explicit reasoning traces before delivering a final judgment, enhancing transparency and interpretability.
  • Dilemma Analysis: For a given (situation, character description, action) triple, it answers two questions: whether the action is acceptable (Yes/No/Ambiguous) and if performing/not performing it would cause emotional/mental discomfort (Yes/No).
  • Self-Generated Training Data: Utilizes self-generated reasoning traces, rejection-sampled against ground truth per character perspective, for robust training.

Good For

  • Applications requiring deep ethical analysis and moral reasoning in the Korean language.
  • Research into character-centric moral decision-making and dilemma resolution.
  • Developing AI systems that need to provide explainable ethical judgments from specific viewpoints.
  • Exploring the complexities of moral reasoning where a single 'correct' answer is not always available, focusing instead on perspective-driven justification.