JinNian0072/efficodebench-inputs

Archived testcases with actual inputs

Each row contains an actual saved stdin input, not a hash or an archive pointer. Parquet is a standard tabular dataset format readable with Hugging Face Datasets, PyArrow or pandas. Select the python or java configuration to view cases.

Python: 31,387 rows / 304 problems / 900 associated pairs. Java: 13,581 rows / 294 problems / 900 associated pairs. C++ testcases are not included. The separate 2700-pair code benchmark is unchanged.

Columns

Column	Meaning
language	Python or Java
problem_id	Problem associated with this testcase
pair_ids	All benchmark pairs sharing this problem's input pool
case_index	Zero-based position in that problem's saved input pool
input	Actual stdin text, including original whitespace and line endings
output	Saved expected-output text; null means unavailable, not empty
input_encoding	utf-8, or base64 only for non-UTF-8 bytes
output_encoding	Same encoding rule; null for unavailable outputs
input_validity	Original constraint-validation status, not upgraded by conversion
output_correctness	Original expected-output certification status

One row is one original testcase entry. Inputs shared by multiple pairs are not copied into separate large rows for every pair. To obtain a pair's inputs, select rows whose pair_ids contains that ID and order them by case_index. Repeated entries within a problem are retained at different indices; do not deduplicate them silently when reproducing a workload.

For a downloaded local dataset directory:

from datasets import load_dataset

cases = load_dataset("parquet", data_files="data/python/test-*.parquet", split="train")
pair_cases = cases.filter(lambda row: "python_14212" in row["pair_ids"])
pair_cases = pair_cases.sort("case_index")
print(pair_cases[0]["input"])

When loading from the Hub, use this repository's ID with configuration python or java and split test. Streaming avoids downloading the entire Java dataset.

For execution, use row["input"].encode("utf-8") when the encoding is utf-8, otherwise base64.b64decode(row["input"]). Apply the same rule to non-null outputs. This reconstructs the original bytes without newline normalization.

Scope and limitations

These are recovered archived input pools, not certified reconstructions of the exact historical measured subsets. Constraint validity and saved Python outputs are not all independently certified. Every Java output remains null: no answers have been invented. Availability is not a new correctness or runtime label certification.

Before new experiments, validate input constraints and correctness checks, freeze the exact selected cases, and record per-testcase measurements. Conversion preserves every input byte, available output byte, problem/pair association, testcase index and validation flag. It does not execute candidates or change the original code benchmark's labels.