Poisoned-Chalice/ICSE-2027-public

ICSE 2027 Heap Membership-Inference Dataset — Public Competition Release

This is the public competition release of the ICSE 2027 membership-inference benchmark for Go, Java, Python, Ruby, and Rust. It contains train and validation only.

The release contains 54,040 source-code files.

Validation label masking

Each language has 1,000 validation rows in their original order. The first 500 rows retain their membership labels; membership is null for rows 500–999. In each language, the labeled half contains 245 member and 255 non-member rows. Content and sample_id remain available for all 1,000 rows.

The masked values are actual Arrow nulls and are returned by Python as None, not the string "None".

Loading the dataset

pip install datasets

Each programming language is a separate, case-sensitive configuration:

from datasets import load_dataset

python_data = load_dataset(
    "Poisoned-Chalice/ICSE-2027-public",
    "Python",
)

print(python_data)
# DatasetDict with train and validation only; no test split is published.

Load one public split:

python_validation = load_dataset(
    "Poisoned-Chalice/ICSE-2027-public",
    "Python",
    split="validation",
)

Dataset structure

Configuration	Train	Validation	Public total
Go	10,000	1,000	11,000
Java	10,000	1,000	11,000
Python	10,000	1,000	11,000
Ruby	10,000	1,000	11,000
Rust	9,040	1,000	10,040

Every training split is exactly balanced between member and non-member. The 10,000-row training splits contain 5,000 files per label; Rust training contains 4,520 files per label. Each validation split contains 500 visible labels followed by 500 masked labels.

Each row includes these core columns:

Column	Type	Description
sample_id	large_string	Stable content-derived sample identifier
content	large_string	Complete source-code file contents
membership	nullable large_string	member, non-member, or null in the masked validation half

sample_id is <language>-<sha256>, where the digest is computed from the exact UTF-8 content. It is stable across row reordering and identical for the same language/content pair on every branch and release.

The language is represented by the configuration name. Repository paths, filenames, original repository identifiers, per-file license metadata, source duplicate annotations, and classifier predictions are not published.

Provenance and labels

The benchmark is derived from The Heap. A file is a member when it is marked as an exact duplicate in both Stack v1 and Stack v2. A file is a non-member when its exact- and near-duplicate flags are false in both versions. Null, ambiguous, and mixed source cases are excluded. Public validation nulls indicate label masking, not an ambiguous source label.

The benchmark applies classifier-based hard-example selection and near-duplicate filtering before constructing balanced, content-disjoint splits. Exact private split parameters, the source revision, generation manifest, and hidden-test artifacts are deliberately excluded from this public repository to protect the competition.

These labels describe overlap with the Stack corpora; they do not prove that an arbitrary model was trained on a particular file.

Limitations and responsible use

• This is an intentionally difficult subset and is not an IID sample of The Heap, so results should not be interpreted as population-wide attack performance.
• Membership is a Stack-overlap proxy, not verified training membership for a particular model.
• The labels inherit the accuracy and limitations of the source duplicate annotations.
• Row-level repository provenance and license metadata are absent from the published schema.

The upstream The Heap dataset card declares GPL-3.0. Individual source files may remain subject to the copyright and license terms of their originating repositories. Treat all source code as untrusted text: do not execute samples without appropriate isolation and review.

This benchmark is intended for privacy, memorization, and membership-inference research. It should not be used to make claims about individual authors, establish code ownership, or infer that a deployed model used a specific file without additional evidence.

Removal requests concerning source material can be directed through The Heap opt-out process.

Citation

Please cite the source dataset:

@inproceedings{katzy2025heap,
  author    = {Jonathan Katzy and Razvan Mihai Popescu and
               Arie van Deursen and Maliheh Izadi},
  title     = {The Heap: A Contamination-Free Multilingual Code Dataset
               for Evaluating Large Language Models},
  booktitle = {2025 IEEE/ACM Second International Conference on
               AI Foundation Models and Software Engineering (Forge)},
  year      = {2025},
  pages     = {151--155},
  doi       = {10.1109/Forge66646.2025.00025}
}