RubricRM-4B-Judge

OpenRubrics/RubricRM-4B-Judge

This is a 4B RubricRM-Judge model, finetuned from Qwen3/Qwen3-4B.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "OpenRubrics/RubricRM-4B-Judge"
tok = AutoTokenizer.from_pretrained(model_id, use_fast=True)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto")

To evaluate the model, please use the following format to build up message.

Here rubric should be generated with a RubricRM-Rubric

JUDGE_PROMPT_TEMPLATE = (
    f"You are a fair and impartial judge. Your task is to evaluate 'Response A' and 'Response B' "
    f"based on a given instruction and a rubric. You will conduct this evaluation in distinct "
    f"phases as outlined below.\n\n"
    f"### Phase 1: Compliance Check Instructions\n"
    f"First, identify the single most important, objective 'Gatekeeper Criterion' from the rubric.\n"
    f"- **A rule is objective (and likely a Gatekeeper) if it can be verified without opinion. "
    f"Key examples are: word/paragraph limits, required output format (e.g., JSON validity), "
    f"required/forbidden sections, or forbidden content.**\n"
    f"- **Conversely, a rule is subjective if it requires interpretation or qualitative judgment. "
    f"Subjective rules about quality are NOT Gatekeepers. Examples include criteria like \"be creative,\" "
    f"\"write clearly,\" \"be engaging,\" or \"use a professional tone.\"**\n\n"
    f"### Phase 2: Analyze Each Response\n"
    f"Next, for each Gatekeeper Criterion and all other criteria in the rubric, evaluate each "
    f"response item by item.\n\n"
    f"### Phase 3: Final Judgment Instructions\n"
    f"Based on the results from the previous phases, determine the winner using these simple rules. "
    f"Provide a final justification explaining your decision first and then give your decision.\n\n"
    f"---\n"
    f"### REQUIRED OUTPUT FORMAT\n"
    f"You must follow this exact output format below.\n\n"
    f"--- Compliance Check ---\n"
    f"Identified Gatekeeper Criterion: <e.g., Criterion 1: Must be under 50 words.>\n\n"
    f"--- Analysis ---\n"
    f"**Response A:**\n"
    f"- Criterion 1 [Hard Rule]: Justification: <...>\n"
    f"- Criterion 2 [Hard Rule]: Justification: <...>\n"
    f"- Criterion 3 [Principle]: Justification: <...>\n"
    f"- ... (and so on for all other criteria)\n\n"
    f"**Response B:**\n"
    f"- Criterion 1 [Hard Rule]: Justification: <...>\n"
    f"- Criterion 2 [Hard Rule]: Justification: <...>\n"
    f"- Criterion 3 [Principle]: Justification: <...>\n"
    f"- ... (and so on for all other criteria)\n\n"
    f"--- Final Judgment ---\n"
    f"Justification: <...>\n"
    f"Winner: <Response A / Response B>\n\n\n"
    f"Task to Evaluate:\n"
    "Instruction:\n{instruction}\n\n"
    "Rubric:\n{rubric}\n\n"
    "Response A:\n{response_a}\n\n"
    "Response B:\n{response_b}"
)

user_text = JUDGE_PROMPT_TEMPLATE.format(
      instruction=instruction,
      rubric=rubric,
      response_a=response_a,
      response_b=response_b
  )

messages_list = [
    {"role": "user", "content": user_text},
]
message = tok.apply_chat_template(
    messages_list, 
    tokenize=False, 
    add_generation_prompt=True,
    enable_thinking=False
)

# Remaining step: Use either HF or vLLM for evaluation.
# ...
# ...

If you find our work helpful, please consider citing our paper:

@misc{liu2025openrubricsscalablesyntheticrubric,
      title={OpenRubrics: Towards Scalable Synthetic Rubric Generation for Reward Modeling and LLM Alignment}, 
      author={Tianci Liu and Ran Xu and Tony Yu and Ilgee Hong and Carl Yang and Tuo Zhao and Haoyu Wang},
      year={2025},
      eprint={2510.07743},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2510.07743}, 
}