README.md

metadata

license: cc-by-4.0
pretty_name: ORBench Decision Problems
task_categories:
  - other
  - text-generation
tags:
  - operations-research
  - model-benchmarking
  - structured-generation
  - synthetic
  - evaluation
language:
  - en
  - fa
size_categories:
  - n<1K

ORBench problems and recorded results

100 synthetic inventory, assignment, and scheduling instances for the ORBench demo.

Organization dataset and collection are public. Live Gradio and the organization card are created by scripts/publish.py. Runnable Space source is stored in demo/. Collection: Aria AI — OR + LLM Evaluation.

Generated data, seed 25. All identifiers are synthetic. Gold objectives come from OR-Tools CP-SAT on the same schema that ships in each row. LLM rows in results.jsonl are explicitly not_evaluated.

Files

• data/problems.jsonl
• data/results.jsonl
• data/leaderboard.csv
• data/splits.json
• data/sample_preview.json
• data/protocol.md
• demo/ — runnable Gradio Space source
• showcase/ — organization static card

Splits

• development: 60 instances, templates D-A / D-B / D-C with a SCHEMA_BEGIN block
• test: 40 instances, templates T-A / T-B in prose
• interactive: 20 ids listed in splits.json

Development and test do not share (family, template_id) pairs. Near-duplicate prompt templates are therefore not reused across the split.

Metrics

Independent validators recompute inventory balance, pack multiples, budget, capacity, assignment uniqueness/eligibility/capacity, and schedule overlap/precedence/release. Optimality gap is reported only when gold status is optimal. An unevaluated method is never written as 0.

License

CC-BY-4.0. Keep the synthetic-data label.