principle_centric_benchmark

Website for the ACL 2026 EvalEval Paper paper "From Rubrics to Recipe: Principle-Centric Benchmark for Evaluating Large Language Models". Dataset and code can be found here.

Data

static/score.csv contains evaluation results for all benchmark instances.

Column	Description
score	Evaluation score assigned to the model output. Scores are on a scale of 0-5, where higher values indicate better performance.
model	The evaluated model. Current models include gemma, gpt, qwen3b, qwen7b, and qwen14b.
task	Task category associated with the example.
principle	Human readable standards or requirements that shape the output.
input	Input prompt provided to the model.
output	Model-generated response.

Citation

@inproceedings{hayati-etal-2026-from,
    title = "From Rubrics to Recipe: Principle-Centric Benchmark for Evaluating Large Language Models",
    author = "Hayati, Shirley Anugrah and Wang, Ruizi and Kang, Dongyeop",
    booktitle = "ACL 2026 Workshop on Evaluating Evaluations (EvalEval)",
    month = jul,
    year = "2026",
    address = "San Diego, United States of America",
    publisher = "Association for Computational Linguistics",
}