R3-Qwen3-4B-14k is part of the R3 family, a series of Robust Rubric-Agnostic Reward Models. We perform SFT on the Qwen3 model family on the 4B, 8B, and 14B scales as well as on Phi-4-reasoning plus. Check out our paper for more information!

Model description

Model type: A reward model trained on a curated R3 dataset collected from 45 diverse sources 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.
Language(s) (NLP): English
License: Apache 2.0
Finetuned from model: Qwen/Qwen3-4B

Model Sources

Project Page: https://rubricreward.github.io
Repository: https://github.com/rubricreward/r3
Paper: https://arxiv.org/abs/2505.13388

Using the Model

python

from transformers import AutoTokenizer
from vllm import LLM, SamplingParams

model_path = "rubricreward/R3-Qwen3-4B-14k"
tokenizer = AutoTokenizer.from_pretrained(model_path)
sampling_params = SamplingParams(temperature=0.6, top_p=0.95, max_tokens=8192, min_p=0, top_k=20)

llm = LLM(
  model=model_path,
  dtype="bfloat16",
  max_model_len=10000,
  tensor_parallel_size=2,
  gpu_memory_utilization=0.9,
  enforce_eager=True,
)

messages: list[dict[str, str]] = [
  {'content': "Evaluate the response based on the given task, input, response, and evaluation rubric. Provide a fair and detailed assessment following the rubric...", 'role': 'user'}
]

list_text = tokenizer.apply_chat_template(
  messages,
  tokenize=False,
  add_generation_prompt=True,
  enable_thinking=True # Switch between thinking and non-thinking modes. 
)

outputs = llm.generate(list_text, sampling_params)

License and use

R3 is licensed under the Apache 2.0 license.

Citation

bibtex

@article{anugraha2025r3,
  title={R3: Robust Rubric-Agnostic Reward Models},
  author={Anugraha, David and Tang, Zilu and Miranda, Lester James V. and Zhao, Hanyang and Farhansyah, Mohammad Rifqi and Kuwanto, Garry and Wijaya, Derry and Winata, Genta Indra},
  journal={arXiv preprint arXiv:2505.13388},
  year={2025}
}