metadata
pretty_name: UlumBench
license: cc-by-4.0
language:
- en
task_categories:
- question-answering
- text-generation
size_categories:
- 1K<n<10K
configs:
- config_name: academic
default: true
data_files:
- split: test
path: academic/evaluation.jsonl
- config_name: diagnostic
data_files:
- split: test
path: diagnostic/families.jsonl
UlumBench
UlumBench contains 6,167 academic questions across 140 topics and 500 constructed multi-step problem families.
Academic questions
- 4,158 multiple-choice questions and 2,009 open-ended questions.
- 743 questions with images.
- Each record includes a question ID, topic, question type, question text, image references, reference answer and evaluation metadata.
Multi-step problems
- 500 families containing 2,400 steps.
- Chain lengths of 2, 4, 6 and 8 steps, with 100, 200, 100 and 100 families, respectively.
- Each family includes its source question IDs, shared context, ordered steps, dependencies, intermediate reference answers, answer types and scoring tolerances.
Supporting data
The dataset also includes question images, curated multiple-choice targets, multipart groupings, source metadata, context variants, a 658-question computer-science panel, and 300 intermediate-state error injections with alternative trajectories and eligibility labels.