AutoRubric

A Python library for evaluating text outputs against weighted criteria using LLM-as-a-judge.

  @misc{rao2026autorubric,
        title={Autorubric: A Unified Framework for Rubric-Based LLM Evaluation},
        author={Delip Rao and Chris Callison-Burch},
        year={2026},
        eprint={2603.00077},
        archivePrefix={arXiv},
        primaryClass={cs.CL},
        url={https://arxiv.org/abs/2603.00077},
  }

Installation

pip install autorubric

Quick Example

import asyncio
from autorubric import Rubric, LLMConfig
from autorubric.graders import CriterionGrader

async def main():
    grader = CriterionGrader(llm_config=LLMConfig(model="openai/gpt-5.1-mini"))

    rubric = Rubric.from_dict([
        {"weight": 10.0, "requirement": "States NMC cell-level energy density in the 250-300 Wh/kg range"},
        {"weight": 8.0, "requirement": "Identifies LFP thermal runaway threshold (~270°C) as higher than NMC (~210°C)"},
        {"weight": 6.0, "requirement": "States LFP cycle life advantage (2000-5000 cycles vs 1000-2000 for NMC)"},
        {"weight": -15.0, "requirement": "Incorrectly claims LFP has higher gravimetric energy density than NMC"}
    ])

    result = await rubric.grade(
        to_grade="""NMC cathodes (LiNixMnyCozO2) achieve 250-280 Wh/kg at the cell level,
        while LFP (LiFePO4) typically reaches 150-205 Wh/kg. However, LFP offers superior
        thermal stability with decomposition onset at ~270°C compared to ~210°C for NMC,
        and delivers 2000-5000 charge cycles versus 1000-2000 for NMC.""",
        grader=grader,
        query="Compare NMC and LFP cathode materials for EV battery applications.",
    )

    print(f"Score: {result.score:.2f}")
    for criterion in result.report:
        print(f"  [{criterion.final_verdict}] {criterion.criterion.requirement}")

asyncio.run(main())

Documentation

Full documentation, API reference, and a cookbook with several dozen recipes are available at autorubric.org.

Features

License

MIT License - see LICENSE file for details.

Acknowledgments

This research was developed with funding from the Defense Advanced Research Projects Agency’s (DARPA) SciFy program (Agreement No. HR00112520300). The views expressed are those of the author and do not reflect the official policy or position of the Department of Defense or the U.S. Government.