---
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

```python
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.