README.md

metadata

license: cc-by-4.0
language:
  - en
pretty_name: Complex Constraints Benchmark Set
size_categories:
  - n<1K
task_categories:
  - text-generation
tags:
  - instruction-following
  - benchmark
  - llm-evaluation
configs:
  - config_name: default
    data_files:
      - split: test
        path: ComplexConstraints_Benchmark_Set.csv

Complex Constraints Benchmark Set

A benchmark for evaluating how well language models follow complex, multi-constraint instructions. It contains 75 items (CIF-001–CIF-075), each a realistic prompt paired with 10–40 evaluation criteria (1,559 total) describing what a correct response must satisfy. Criteria are meant for rubric-based grading (human or LLM-as-a-judge), not exact match.

Structure

Single wide-format CSV, one item per row:

• benchmark_id, prompt, use_case, instruction_type, prompt_style
• For each criterion i (1–40): criterion_{i}. Unused criterion columns are empty.

Loading

from datasets import load_dataset
ds = load_dataset("USERNAME/complex-constraints-benchmark", split="test")  # TODO: set repo id
print(ds[0]["prompt"])

Or with pandas:

import pandas as pd
df = pd.read_csv("ComplexConstraints_Benchmark_Set.csv")

License & Citation

Released under CC-BY-4.0. You may use, redistribute, and adapt it for any purpose, including commercially, as long as you give appropriate credit.

@misc{complex_constraints_benchmark,
  title  = {Complex Constraints Benchmark Set},
  author = {TODO: authors},
  year   = {2025},
  url    = {TODO: dataset URL}
}