ReVeR-Traces-Judge

Supervised fine-tuning data for the agentic judge of ReVeR. Each of the 60,620 trajectories scores one response against every item of a rubric. The judge may call a Python tool for mechanical checks such as counting words; 58.9% of the trajectories contain a tool call.

The companion dataset ReVeR-Traces-Rubric holds the rubric-generation examples for the same rubrics. The two datasets are aligned by id: every rubric there is used, word for word, by exactly two trajectories here, one for each response of a preference pair.

How the data was built

1. For each preference pair (an instruction, a preferred response and a rejected response), a teacher model (DeepSeek-V4-Flash with reasoning) wrote a rubric from the instruction alone. Every rubric item is labelled [Hard Rule] or [Principle].
2. The same teacher judged each response separately against that rubric, without seeing the other response: 0 or 1 for a Hard Rule, an integer from 0 to 100 for a Principle. It could make at most one Python tool call per response.
3. A pair is kept only if the judge scores rank the preferred response strictly higher. The score of a response is the number of satisfied Hard Rules plus the mean Principle score divided by 101, so Principles only break ties between responses that satisfy the same number of Hard Rules.

Fields

Field	Content
id	Id of the preference pair. It joins the two trajectories of a pair to their rubric in ReVeR-Traces-Rubric
source	Origin of the pair: helpsteer3, ultrafeedback, skywork, magpie or ifeval (synthetic instruction-following prompts)
candidate	chosen for the preferred response of the pair, rejected for the other one
messages	The trajectory, described below
tools	The definition of the Python tool offered to the judge, as a JSON string

messages holds, in order:

• system: the judge prompt.
• user: a JSON object with instruction, response and rubrics (each item has a rubric_id and a criterion).
• assistant: the judge's reasoning in <think>…</think>. If it uses the tool, a tool_call message and a tool_response message follow, then another assistant message.
• The last assistant message ends with a <judge> block: a JSON object with one score and one explanation per rubric item.

Usage

import json
from datasets import load_dataset

ds = load_dataset("OpenRubrics/ReVeR-Traces-Judge", split="train")

example = ds[0]
task = json.loads(example["messages"][1]["content"])
instruction, response, rubrics = task["instruction"], task["response"], task["rubrics"]
final = example["messages"][-1]["content"]

Citation

If you use this dataset, please cite our paper Rubrics from the Task, Verdicts from Evidence: Pointwise Rubric Rewards with Tool-Augmented Judges:

@misc{liu2026rubricstaskverdictsevidence,
      title={Rubrics from the Task, Verdicts from Evidence: Pointwise Rubric Rewards with Tool-Augmented Judges}, 
      author={Tianci Liu and Haoyu Wang},
      year={2026},
      eprint={2610.18766},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2610.18766}, 
}