README.md

metadata

pretty_name: Program Pair Data
configs:
  - config_name: cpp
    data_files:
      - split: test
        path: final_benchmark_uniform/cpp_seed42_sanitized.jsonl
  - config_name: python
    data_files:
      - split: test
        path: final_benchmark_uniform/python_seed42_sanitized.jsonl
  - config_name: java
    data_files:
      - split: test
        path: final_benchmark_uniform/java_seed42_sanitized_process.jsonl
task_categories:
  - text-classification
language:
  - en
tags:
  - code
  - efficiency
  - benchmark
size_categories:
  - 1K<n<10K

Program Pair Data

This dataset contains 2,700 main-evaluation program pairs. Each language has one test split with 900 pairs.

Configuration	Pairs	Problems	Maximum pairs per problem
cpp	900	900	1
python	900	221	5
java	900	320	5

Each language contains 300 pairs in each slow-to-fast execution-cost ratio range: [2, 4), [4, 8), and [8, infinity).

Fields

• language, problem_id, pair_id: language, task and comparison identifiers.
• slow_code, fast_code: exact model-facing program text, ordered by the saved efficiency labels.
• slow_code_id, fast_code_id: SHA-256 hashes of the displayed source text.
• measured_slow_code_id, measured_fast_code_id: SHA-256 identifiers of the measured source programs, which can differ from the displayed source.
• slow_id, fast_id: measured implementation identifiers.
• speedup: the saved slow-to-fast execution-cost ratio.
• measurement_metric: PIE released gem5 runtime for C++, CPU instruction count for Python, or elapsed execution time for Java.
• stratum, sampling_key: ratio stratum and deterministic sampling key.
• slow_time, fast_time: released aggregate costs, present for C++.

The Python and Java ratios aggregate per-input measurement ratios; they are not reconstructed as ratios of aggregate times. Test inputs, expected outputs, per-run measurements and original measured-source text are not included. This dataset provides the program pairs and saved efficiency labels used in the main evaluation.

Loading

from datasets import load_dataset

data = load_dataset("JinNian0072/program-pair-data", "python", split="test")
assert len(data) == 900
print(data[0]["slow_code"], data[0]["fast_code"])

Use cpp or java for the other languages. The JSONL files are uncompressed. RELEASE.json contains record counts and file checksums. Upstream code and data retain their respective terms.