metadata
license: apache-2.0
language:
- aa
- af
- ar
- as
- az
- be
- bg
- bn
- bs
- ca
- cs
- da
- de
- el
- en
- es
- et
- eu
- fa
- fi
- fr
- ha
- he
- hi
- hr
- hu
- hy
- id
- ie
- it
- iw
- ja
- ka
- kk
- ko
- ku
- la
- lt
- lv
- mk
- ms
- my
- nl
- nn
- 'no'
- oc
- pl
- pt
- ro
- ru
- rw
- sa
- sco
- si
- sk
- sl
- sr
- sv
- sw
- ta
- th
- tl
- tlh
- tr
- tt
- uk
- vi
- vo
- war
- xh
- zh
datasets:
- rubricreward/mR3-Dataset-100K-EasyToHard
base_model:
- Qwen/Qwen3-8B
pipeline_tag: text-generation
library_name: transformers

mR3-Qwen3-8B-tgt-prompt-tgt-thinking
mR3-Qwen3-8B-tgt-prompt-tgt-thinking is part of the mR3 family, a series of Multilingual Rubric-Agnostic Reward Reasoning Models. We perform SFT on the Qwen3 model family on the 4B, 8B, and 14B scales. Check out our paper for more information!
Model description
- Model type: A reward model trained on a curated mR3 dataset collected from 72 languages that covers tasks such as classification, preference optimization, and question answering. Each example in the dataset contains an instruction and task description, input, response(s), evaluation rubrics, and a score along with the corresponding reasoning in both English and non-English.
- Number of Language(s) (NLP): 72 languages
- License: Apache 2.0
- Finetuned from model: Qwen/Qwen3-8B
Model Sources
- Project Page: https://rubricreward.github.io
- Repository: https://github.com/rubricreward/mr3
- Paper: https://arxiv.org/abs/2510.01146
Using the Model
For the following examples, we will use messages
as our pairwise task example.
Click to reveal the example prompt
system_prompt = """# μ§μ
κ·νλ 곡μ ν μ¬μ¬μμμΌλ‘μ, μ£Όμ΄μ§ μ¬μ©μ ν둬ννΈμ λν΄ λ κ°μ 보쑰 λ΅λ³ μ€ μ΄λ κ²μ΄ λ λμμ§ νκ°ν΄μΌ ν©λλ€. λ€μ κΈ°μ€μ μμμ λ°λΌ 체κ³μ μΌλ‘ λ΅λ³μ λΉκ΅νμΈμ
# νκ°κΈ°μ€
Assistant B: Assistant Bλ μ λ°μ μΌλ‘ λ λμ μλ΅μ μ 곡ν©λλ€. νλ¨ κΈ°μ€μ λ€μ μμλλ‘ μ
λλ€ β μμ μ± λ° μ μ μ±, μ μ©μ±, κ΄λ ¨μ±, κ°κ²°μ±, μ μ€ν¨, κ·Έλ¦¬κ³ ν¬κ΄μ±μ
λλ€.
Assistant A: Assistant Aλ μ λ°μ μΌλ‘ λ λμ μλ΅μ μ 곡ν©λλ€. νλ¨ κΈ°μ€μ λ€μ μμλλ‘ μ
λλ€ β μμ μ± λ° μ μ μ±, μ μ©μ±, κ΄λ ¨μ±, κ°κ²°μ±, μ μ€ν¨, κ·Έλ¦¬κ³ ν¬κ΄μ±μ
λλ€.
# μλ΅ νμ
{'type': 'object', 'properties': {'explanation': {'type': 'string', 'description': 'νκ° μμλ₯Ό κΈ°λ°μΌλ‘ Assistant Aμ Assistant Bλ₯Ό λΉκ΅ν κ°λ¨ν μ΄μ μ€λͺ
μ
λλ€.'}, 'score': {'type': 'string', 'description': "νμ κ²°κ³Ό: 'Assistant A' λλ 'Assistant B' μ€ νλμ
λλ€.", 'enum': ['Assistant A', 'Assistant B']}}, 'required': ['explanation', 'score']}
"""
user_prompt = """# μ
λ ₯(μ¬μ©μ ν둬ννΈ)
μ²νμ΄λ 무μμΈκ°?
# Assistant A
μ²ν"μ λν μ μλ₯Ό μ 곡ν μ μμ΅λλ€. λ§₯λ½μ λ°λΌ μλ―Έκ° λ€λ₯Ό μ μκΈ° λλ¬Έμ
λλ€. κ·Έλ¬λ λ²μ μ²λ²μ΄λ νλ ₯ νμμ΄λ μ΄λ€ ννμ μ²νμ΄λ κ°λ±μ ν΄κ²°νκ±°λ μ¬ν λ¬Έμ λ₯Ό ν΄κ²°νλ λ° μ ν©ν ν΄κ²°μ±
μ΄ μλλΌλ μ μ κ°μ‘°νκ³ μΆμ΅λλ€. λͺ¨λ μ¬λμ μλͺ
, μ‘΄μμ±, μ‘΄μ€μ λ°μ κΆλ¦¬κ° μμΌλ©°, μ°¨μ΄μ κ°λ±μ ν΄κ²°νλ ννλ‘κ³ κ±΄μ€μ μΈ λ°©λ²μ μ°Ύλ κ²μ΄ μ€μν©λλ€. λ€λ₯Έ μ§λ¬Έμ΄λ μ°λ € μ¬νμ΄ μμΌλ©΄ μΈμ λ μ§ λ¬ΈμνμΈμ. λμμ΄ λκ³ μμ ν λ΅λ³μ μ 곡νκΈ° μν΄ μ΅μ μ λ€νκ² μ΅λλ€.
# Assistant B
μ¬ν μ§νμ λ²μμ΄ νλ, μΌλ°μ μΌλ‘ νμ¬ μ μ£ νκ²°μ λν μ§μνμ μ§ννλλ‘ λͺ
λ Ήνλ λ²μ μ μ°¨μ
λλ€. μ¬κΈ°μλ κ΄ν κΆμ λ²λ₯ μ λ°λΌ μΉλͺ
μ μ£Όμ¬, κ°μ μ¬, κ΅μν λλ λ€λ₯Έ μ¬ν λ°©λ²μ΄ ν¬ν¨λ μ μμ΅λλ€. μ¬νμ΄ μλ κ²½μ°, μ¬ν μ§νμ μ§μνμ μ§ννλ κ²μ μλ―Ένλ©°, μ¬κΈ°μλ ꡬκΈ, 보νΈκ΄μ°° λλ κΈ°ν μ²λ²μ΄ ν¬ν¨λ μ μμ΅λλ€.
# λΉμ μ μλ΅
"""
# prepare the model input
messages = [
{'role': 'system', 'content': system_prompt},
{'role': 'user', 'content': user_prompt}
]
π§ Using transformers
Below is an example of using our mR3-Qwen3-8B-tgt-prompt-tgt-thinking
model by using an non-English prompt and a non-English reasoning using language forcing and π€ transformers
:
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "rubricreward/mR3-Qwen3-8B-tgt-prompt-tgt-thinking"
# Load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype="auto",
device_map="auto"
)
text = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
enable_thinking=True # Switches between thinking and non-thinking modes. Default is True.
)
### Key difference: Language forcing in Korean ###
text += "μκ² μ΅λλ€. μ λ μ 곡λ λͺ¨λ μ 보λ₯Ό μ μ€νκ² κ²ν νκ³ μ£Όμ΄μ§ νκ° κΈ°μ€μ λ°λΌ νκ°ν λ€, μμ²λ νμμ λ§μΆ° μ λ΅λ³μ νκ΅μ΄λ‘ λͺ
ννκ² μκ°νλ©° μ μνκ² μ΅λλ€."
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
# Conduct text completion
generated_ids = model.generate(
**model_inputs,
max_new_tokens=16384,
temperature=0.6, top_p=0.95, min_p=0, top_k=20
)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()
# Parsing thinking content
try:
# rindex finding 151668 (</think>)
index = len(output_ids) - output_ids[::-1].index(151668)
except ValueError:
index = 0
content = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip("\n")
print(content)
β‘ Using vLLM
Alternatively, you may also use vLLM
for faster inference by including language forcing:
from transformers import AutoTokenizer
from vllm import LLM, SamplingParams
model_path = "rubricreward/mR3-Qwen3-8B-tgt-prompt-tgt-thinking"
tokenizer = AutoTokenizer.from_pretrained(model_path)
sampling_params = SamplingParams(temperature=0.6, top_p=0.95, max_tokens=16384, min_p=0, top_k=20)
llm = LLM(
model=model_path,
dtype="bfloat16",
max_model_len=32768,
)
list_text = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
enable_thinking=True # Switch between thinking and non-thinking modes.
)
for index in range(len(list_text)):
### Key difference: Language forcing in Korean ###
list_text[index] += "μκ² μ΅λλ€. μ λ μ 곡λ λͺ¨λ μ 보λ₯Ό μ μ€νκ² κ²ν νκ³ μ£Όμ΄μ§ νκ° κΈ°μ€μ λ°λΌ νκ°ν λ€, μμ²λ νμμ λ§μΆ° μ λ΅λ³μ νκ΅μ΄λ‘ λͺ
ννκ² μκ°νλ©° μ μνκ² μ΅λλ€."
outputs = llm.generate(list_text, sampling_params)
print(outputs[0].output.text)
License and use
mR3 is licensed under the Apache 2.0 license.
Citation
@article{anugraha2025mr3,
title={mR3: Multilingual Rubric-Agnostic Reward Reasoning Models},
author={Anugraha, David and Hung, Shou-Yi and Tang, Zilu and Lee, Annie En-Shiun and Wijaya, Derry and Winata, Genta Indra},
journal={arXiv preprint arXiv:2510.01146},
year={2025}
}