---
license: apache-2.0
library_name: transformers
base_model: Qwen/Qwen3-4B
pipeline_tag: text-generation
tags:
  - moral reasoning
  - value reasoning
  - persona
  - chain-of-thought
language:
  - ko
---

# Model Card for MET-D-Qwen3-4B-ko-only

MET-D-Qwen3-4B-ko-only is a Korean-only moral reasoning model fine-tuned from [Qwen3-4B](https://huggingface.co/Qwen/Qwen3-4B). Given a moral dilemma, a character description, and a candidate action, it judges the action from that character's perspective and explains its judgment with an explicit chain-of-thought before answering. Moral dilemmas rarely have a single correct answer, which makes reasoning traces hard to verify. We address this by introducing a character perspective that yields a ground-truth answer, which is used for rejection-sampling the model's own reasoning traces, conditioned on a per-language, per-situation selection of theoretical grounds. Both the reasoning trace and the final answer are generated in Korean.

## Model Details

- **Base model:** Qwen/Qwen3-4B
- **Task:** for a given `(situation, character description, action)` triple, answer two questions from the character's perspective:
  1. Is the action acceptable? (`Yes` / `No` / `Ambiguous`)
  2. Would (not) doing it be emotionally/mentally discomforting? (`Yes` / `No`)
- **Training data:** self-generated reasoning traces, rejection-sampled against the ground truth per character perspective, in Korean only
- **Language coverage:** this checkpoint is trained on **Korean only**. The multilingual model and other single-language variants are also available (see [Model Variants](#model-variants) below).
- **License:** Apache 2.0 (inherited from the base model)

## How to Get Started

```python
from transformers import AutoTokenizer
from vllm import LLM, SamplingParams

model_id = "launch/MET-D-Qwen3-4B-ko-only"
tokenizer = AutoTokenizer.from_pretrained(model_id)
llm = LLM(model=model_id, max_model_len=16384)

situation = "..."            # the moral dilemma narrative
character = "..."            # the character description
action = "..."                # the candidate action to judge

prompt = f"""도덕적 딜레마와 캐릭터 설명이 주어지며, 당신은 이에 따라 두 가지 질문에 답변해야 합니다.
첫 번째 질문에 대한 대답은 "예", "아니오", 또는 "모호함" 중 하나여야 하며,
두 번째 질문에 대한 대답은 "예" 또는 "아니오" 중 하나여야 합니다.
출력 형식을 엄격히 지켜야 하며, 이는 자동으로 파싱될 예정입니다.
모든 질문에 대해 반드시 "이유:"과 "답:"를 모두 출력하십시오.

도덕적 딜레마:
{situation}

캐릭터 설명:
{character}

질문:
1. A의 관점에서, 다음 행동 — 이미 실행되었거나, 실행을 고려 중이거나, 고려될 수 있거나, 상황 속에서 의문시되고 있는 행동 — 은 용납될 수 있는가? "{action}"
2. 이전 질문에 대한 대답이 '예'인 경우, A는 그 행동을 하는 것에 대해 정신적 또는 감정적으로 불편함을 느낄 것인가? 반대로, 대답이 '아니오'인 경우, A는 그 행동을 하지 않는 것에 대해 정신적 또는 감정적으로 불편함을 느낄 것인가?

당신의 대답:
1. 이유: {{이유}} 답: {{예/아니오/모호함}}
2. 이유: {{이유}} 답: {{예/아니오}}
"""

chat_prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": prompt}],
    tokenize=False,
    add_generation_prompt=True,
)

sampling_params = SamplingParams(temperature=0.0, max_tokens=2048)
outputs = llm.generate(chat_prompt, sampling_params)
print(outputs[0].outputs[0].text)
```

## Model Variants

This checkpoint is part of the [MET collection](https://huggingface.co/collections/launch/met), which includes the same task across base models and language subsets:

| Repo | Base model | Language(s) |
|---|---|---|
| `launch/MET-D-Qwen3-4B` | Qwen3-4B | all 6 (mixed) |
| `launch/MET-D-Qwen3-4B-en-only` | Qwen3-4B | English only |
| `launch/MET-D-Qwen3-4B-es-only` | Qwen3-4B | Spanish only |
| `launch/MET-D-Qwen3-4B-hi-only` | Qwen3-4B | Hindi only |
| `launch/MET-D-Qwen3-4B-ko-only` | Qwen3-4B | Korean only |
| `launch/MET-D-Qwen3-4B-ms-only` | Qwen3-4B | Malay only |
| `launch/MET-D-Qwen3-4B-zh-only` | Qwen3-4B | Chinese only |
| `launch/MET-D-Qwen3-8B` | Qwen3-8B | all 6 (mixed) |
| `launch/MET-D-Qwen3-8B-en-only` | Qwen3-8B | English only |
| `launch/MET-D-Gemma3-4B` | Gemma-3-4B-it | all 6 (mixed) |
| `launch/MET-D-Gemma3-4B-en-only` | Gemma-3-4B-it | English only |

## Citation

If you use this, please cite:

```bibtex
@article{lee2026met,
  title={MET: Theory-Grounded and Culture-Aware Multilingual Moral Reasoning},
  author={Lee, Ayoung and Kwon, Ryan and Zhang, Yunxiang and Liu, Yuxuan and Railton, Peter and Wang, Lu},
  journal={arXiv preprint arXiv:2607.11736},
  year={2026}
}
```