MET-D-Gemma3-4B is a multilingual moral reasoning model fine-tuned from Gemma-3-4B-it. 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 whichever language the prompt is written in, so non-English speakers get legible reasoning in their own language.

Model Details

Base model: google/gemma-3-4b-it
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, covering six languages (English, Spanish, Hindi, Korean, Malay, Chinese)
Language coverage: this checkpoint is trained on all 6 languages combined. Single-language variants are also available (see Model Variants below).
License: Gemma Terms of Use (inherited from the base model)

How to Get Started

python

from transformers import AutoTokenizer
from vllm import LLM, SamplingParams

model_id = "launch/MET-D-Gemma3-4B"
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"""I will provide you a moral dilemma and present you with a character description and questions.
The answer for the first question should be Yes, No, or Ambiguous, and the answer for the second question should be Yes or No.
You should strictly adhere to the output format, since it would be automatically parsed.
Make sure to output both the "Reason:" and "Answer:" for all 2 questions.

Situation:
{situation}

Character Description:
{character}

Question:
1. From A's perspective, would the following action be deemed acceptable? "{action}"
2. If the previous answer is yes, would A find doing the action mentally or emotionally discomforting? Conversely, if the previous answer is no, would A find not doing the action discomforting?

Your Answer:
1. Reason: {{Reason}} Answer: {{Yes/No/Ambiguous}}
2. Reason: {{Reason}} Answer: {{Yes/No}}
"""

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, 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}
}