suchirsalhan/

rubric-adherence

One row per exam question = the question + its complete, authentic marking rubric, taken verbatim from official examination-board mark schemes. Built for controlled interpretability experiments on:

Which components of an authentic human rubric causally control an LLM's grading?

from datasets import load_dataset ds = load_dataset("suchirsalhan/rubric-adherence", split="train") ds[0]["question"], ds[0]["rubric_verbatim"]… See the full description on the dataset page: https://huggingface.co/datasets/suchirsalhan/rubric-adherence.