WONI-Correctness

WONI — Watermark Observatory for Networked Integrity is a collection of benchmark inputs and evaluation artifacts for language-model watermarking research.

Author: Euisuh Jeong
Affiliation: Artificial Intelligence Based Technology Company, Republic of Korea Air Force

Existing code-candidate evaluation labels and diversity audits. Release 2026-09-23-public-v3 (public-v3).

The released held-out files contain 2,112 evaluation records: 132 tasks (HumanEval/32 through HumanEval/163), eight samples per task, two released model archives. The former 2,624 count included development evaluations absent from this repository. Sources: the two heldout-*.eval_results.json files.

Generated solution strings are withheld pending a clear redistribution basis for those third-party release assets. Each retained record provides task_id, sample_index, solution_sha256, base_status and plus_status. Task-local sample order and all pass/fail labels are preserved. Original file hashes and transformations are recorded in PUBLICATION.json.

The two heldout-*.audit.json files retain candidate-pool statistics; the two heldout-*.summary.json files retain fixed-K curves and task-bootstrap intervals. The frontier PNG/SVG and progress logs are historical observations. These metrics do not establish semantic equivalence, watermark capacity or detector TPR/FPR. Sources: each audit's note, each summary's warning.

Source programs came from EvalPlus v0.1.0 release assets: chatgpt_temp_0.8.zip (SHA-256 d30c99aae23e4a8ed9e7ca23c4f25735237c8de7f37a5c02272a8f45da7b68f7) and starcoder_temp_0.8.zip (SHA-256 ceecd2550884320b7523f992004e615bb2dc00ad378aae1555e8988ad88f7a47). The HumanEval_<task>/<sample>.py members identify task/sample source programs; match retained solution hashes before reuse. Program text and tests are not included. Evaluation provenance records EvalPlus 0.2.1. HumanEval's source and terms: OpenAI human-eval. Downloaded code must never be executed directly on a workstation.

Download and integrity

This is an artifact collection, not a homogeneous training table. Download explicit files or a pinned snapshot:

from huggingface_hub import snapshot_download
root = snapshot_download("euisuh/WONI-correctness", repo_type="dataset", revision="public-v3")

Resolve the tag once and record its full Hub commit SHA. Verify payload hashes against MANIFEST.json. Exact inventory and counts: dataset_info.json. This JSON is descriptive project metadata, not a Hugging Face DatasetInfo schema or an automatic load_dataset config.

Terms and related artifacts

NOTICE.md distinguishes original observations from source-derived content. LICENSE preserves MIT for original observations and release documentation only; it does not cover third-party passages, code or model assets. Use the release revision cited in the paper for reproducibility.

Related resources: WONI-CrossBench, WONI-correctness, WONI-robustness, WONI-crossScheme. The detector paper's public replay artifact is v1.2.0. These author-linked repositories are not anonymous review artifacts or claims of paper acceptance. Cite the applicable upstream datasets as well.

Citation

@misc{jeong2026woni_correctness,
  author = {Euisuh Jeong},
  title = {WONI-Correctness},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/euisuh/WONI-correctness},
  note = {Public release 2026-09-23-public-v3; record the resolved commit SHA}
}