Abhisek987/deplab-dependency-compatibility

DepLab Dataset v1.0.0: Empirical Python Dependency Compatibility Experiments

24,922 real-world experiments testing whether pairs of Python packages can be installed and used together. Each experiment records the full outcome across three stages: dependency resolution, import, and a runtime smoke test.

This dataset was built to answer a simple question: how often does an environment that resolves successfully still fail in practice? In the development split, 3,049 environments (14.2%) passed resolution but failed at import or smoke-test time. Resolvers cannot see these failures because they only read published metadata and never install or run anything.

The dataset powers DepLab, an ML-guided tool that predicts working dependency combinations.

Contents

File	Records	Description
deplab-development-results-v1.0.0.jsonl	21,490	Full experiment results, development split
deplab-validation-results-v1.0.0.jsonl	3,432	Full experiment results, sealed validation split
deplab-development-features-v1.0.0.csv	21,490	Model-ready engineered features (142 columns) for the development split

Loading with the datasets library

from datasets import load_dataset

# Full experiment results (both splits)
experiments = load_dataset(
    "Abhisek987/deplab-dependency-compatibility", "experiments")

# Model-ready engineered features
features = load_dataset(
    "Abhisek987/deplab-dependency-compatibility", "features")

Experiment design

Each experiment installs a pair of Python packages at specific versions into a fresh environment and tests them in three sequential stages:

1. Resolution: can a consistent dependency set be produced from published constraints?
2. Import: after installation, do both packages import successfully?
3. Smoke test: do basic interoperability calls between the packages run without error?

The experiment stops at the first failing stage, giving one of four outcomes.

Coverage: 41 unique packages in this release: 35 packages (36 package-pair families) in the development split and 6 completely different packages (4 families) in the sealed validation split. Python 3.8 through 3.14, Linux x86_64, multiple versions per package. A further 9 packages are reserved for a future final test set and are not included in this release.

Outcome distribution

Development split (21,490 experiments):

Outcome	Count	Share
pass	12,031	56.0%
resolution_failure	6,410	29.8%
import_failure	2,448	11.4%
smoke_test_failure	601	2.8%

Sealed validation split (3,432 experiments):

Outcome	Count
resolution_failure	2,714
pass	639
import_failure	60
smoke_test_failure	19

The validation split uses six packages that do not appear anywhere in the development split: boto3, botocore, s3transfer, celery, kombu, and billiard. Validation therefore measures generalization to entirely unseen packages, not just unseen versions of training packages. The split was kept sealed during model development.

JSONL record schema

Each line is one experiment (record schema version 1.3.0) with these top-level fields:

• experiment_id: unique hexadecimal identifier for the experiment
• schema_version: version of the record schema (1.3.0)
• spec: the experiment specification: package_a and package_b (each with name and version), python_version, os, and architecture
• outcome: one of pass, resolution_failure, import_failure, smoke_test_failure
• stages: list of executed stages, each with the exact command, stage name, exit_code, duration_seconds, peak_rss_bytes, captured stdout and stderr, and a timed_out flag
• normalized_error: normalized error text for failed experiments (resolver output for resolution failures, Python traceback for import and smoke-test failures); null on pass
• exception_type: Python exception class (e.g. ValueError, AttributeError) for import and smoke-test failures; null otherwise
• installed_environment: the fully resolved installed package set as name==version strings; empty for resolution failures
• wheel_artifacts: candidate wheels considered for the package pair, with filename, size, sha256, python/abi/platform tags, upload date, yanked status, native-extension flag, and the compatibility decision with its reason
• installed_wheel_artifacts: the wheels actually installed, with filename, package, version, size, sha256, tags, and source URL; empty for resolution failures
• artifact_lock_format / artifact_lock_sha256: lockfile format (pylock.toml, PEP 751) and its SHA-256 hash for reproducibility; null for resolution failures
• runtime: execution environment: OS, architecture, kernel, libc, Python implementation and exact version, and the uv version used
• resources: resource measurements around the run: package cache state before and after, disk, network, and peak memory
• duration_seconds, retry_count, started_at (ISO 8601 UTC), measured: execution metadata

Local filesystem paths from the experiment machines have been replaced with the placeholders <WORKDIR> and <venv>.

Feature CSV

deplab-development-features-v1.0.0.csv contains 142 engineered features per experiment, including:

• Version features: major/minor/patch components, version rank and percentile, release age, release-date distance between the two packages
• Constraint features: declared requirements between the pair, upper and lower bounds, exact pins, whether published constraints allow the combination
• Wheel features: eligible wheel counts, Python/ABI/platform tags and whether they match, wheel sizes, native-extension flags
• Changelog features: mined signals from release notes for both packages, including counts and flags for breaking changes, removals, deprecations, API and ABI changes, dependency changes, and Python support drops
• Labels: outcome, compatibility_label, is_compatible, is_failure, failure_stage

Known limitations

• Experiments test pairs of packages, not full requirements files with many simultaneous pins.
• Coverage is 41 packages in this release; conclusions may not transfer to arbitrary packages.
• All experiments ran on Linux x86_64. Other platforms may behave differently.
• Smoke tests exercise basic interoperability, not full API surfaces.

License

Creative Commons Attribution 4.0 International (CC BY 4.0). You may use, share, and adapt this dataset with attribution.

Citation

If you use this dataset, please cite:

Behera, Abhisek (2026). DepLab Dataset v1.0.0: Empirical Python Dependency
Compatibility Experiments. Zenodo. DOI: 10.5281/zenodo.21729353
https://doi.org/10.5281/zenodo.21729353

Links

• Zenodo (DOI): https://doi.org/10.5281/zenodo.21729353
• Project: https://github.com/Abhisek12378/DepLab
• Live application: https://deplab.13-234-114-139.sslip.io/
• Author: Abhisek Behera
  • GitHub: https://github.com/Abhisek12378
  • Hugging Face: https://huggingface.co/Abhisek987