| ---
|
| license: cc-by-4.0
|
| pretty_name: PyPI API Drift
|
| language:
|
| - en
|
| tags:
|
| - code
|
| - python
|
| - api-evolution
|
| - breaking-changes
|
| - benchmark
|
| - llm-evaluation
|
| task_categories:
|
| - text-classification
|
| size_categories:
|
| - 10M<n<100M
|
| configs:
|
| - config_name: versionprobe
|
| data_files: benchmark/versionprobe.parquet
|
| - config_name: versionprobe_verified
|
| data_files: benchmark/versionprobe_verified.parquet
|
| default: true
|
| - config_name: versionprobe_lite
|
| data_files: benchmark/versionprobe_lite.parquet
|
| - config_name: changes
|
| data_files: data/changes.parquet
|
| - config_name: api_history
|
| data_files: data/api_history.parquet
|
| - config_name: releases
|
| data_files: data/releases.parquet
|
| - config_name: signatures
|
| data_files: data/signatures.parquet
|
| - config_name: module_coverage
|
| data_files: data/module_coverage.parquet
|
| ---
|
|
|
| # PyPI API Drift
|
|
|
| Historical public-API snapshots for up to 80 recent final releases of each of the 1,872 most-downloaded
|
| Python libraries, drawn from 2021 to October 2026 (55,565 releases), and every change between consecutive
|
| releases (15,258,652 events, 210,535 distinct breaking changes, 181,023
|
| high-confidence). It is a high-precision, low-recall record, not complete ground truth: names created at runtime are
|
| often missing. This is a research release; all validation was done by pipeline-isolated AI agents, not human
|
| annotators (a 200-item human labelling kit is in the repository).
|
|
|
| How it was built: wheels are read with HTTP range requests (only the ZIP directory and changed `.py`/`.pyi` files,
|
| about 8.5% of the wheels' bytes, roughly 26 GB instead of 300 GB) and parsed with `ast` without
|
| importing or executing any package code. A name is labelled `removed` only when it is no longer reachable through any
|
| import path, or when a module `__getattr__` explicitly rejects it; names merely dropped from `__all__` are
|
| `left_public_api`. Filter `changes` on `confidence == "high"` and `representative`, and `api_history` on
|
| `coverage_status == "static"`, for the strictest view.
|
|
|
| | Measured on samples the pipeline never saw | Result |
|
| |---|---|
|
| | Precision, high-confidence breaking changes, re-audited on v1.2 | 91.9% (95% CI 86.6–95.2%; n=160) |
|
| | Precision, high-confidence breaking changes, first untuned audit | 88.8% (95% CI 80.0–94.0%; n=80) |
|
| | VersionProbe labels correct (fresh sample) | 91.5% (95% CI 88.4–93.9%; n=400) |
|
| | Recall within the dataset's static public-API definition | 14/16 |
|
| | Recall over all runtime-visible changes | 16/99 (runtime-generated APIs are invisible to static analysis) |
|
|
|
| **VersionProbe** (10,958 pairs from 583 libraries; `versionprobe_lite`: 983 pairs)
|
| asks a model whether a snippet is valid for an exact library version. Each pair is valid in one release and invalid in
|
| the next, and carries the release date so you can split results at a model's training cutoff.
|
| **`versionprobe_verified`** (492 pairs, 984 items, 375 libraries) contains only pairs whose
|
| labels were confirmed at runtime by AI agents isolated from the pipeline (no human adjudication; evidence in the
|
| repository); it is the default configuration and the one to rank models on.
|
|
|
| | Model, closed book, on `versionprobe_verified` | Item accuracy | Pair accuracy |
|
| |---|---|---|
|
| | `claude-opus-5-5` | 58.4% | 26.6% |
|
| | `claude-fable-5-1` | 56.5% | 22.4% |
|
| | `claude-sonnet-5-5` | 54.3% | 18.3% |
|
| | `claude-haiku-5-5` | 49.6% | 9.2% |
|
|
|
| A constant answer scores 50% / 0%. Given API history retrieved from this dataset, `claude-opus-5-5` scored
|
| 100.0%; that is an oracle condition showing the answer is in the record, not a general benefit.
|
|
|
| In a preregistered code-repair experiment on 143 real breaking changes, repeated across 4 models from
|
| one model family, API
|
| history from this dataset raised API-compatible, intent-preserving migrations (under the evaluation protocol: runtime
|
| check of names and argument binding plus an AI judge of intent; not a behavioural test) to 68.5%–69.9%
|
| for every model, from 7.0%–56.6% unaided and 46.2%–62.2% with the new release's
|
| documentation.
|
|
|
| `module_coverage` and the `coverage_status` column of `api_history` mark where static analysis cannot see the whole
|
| API (stub-only, dynamic or compiled modules), so absence of a name is not read as proof that it does not exist.
|
|
|
| ```python
|
| from datasets import load_dataset
|
| probe = load_dataset("pushkar87dg/pypi-api-drift", "versionprobe_verified", split="train")
|
| changes = load_dataset("pushkar87dg/pypi-api-drift", "changes", split="train")
|
| ```
|
|
|
| Use `breaking & ~parent_removed & representative & confidence == "high"` for one row per real, high-confidence
|
| breaking change. Before use, check downloaded tables against `DATA_MANIFEST.json` (data version 2026.10-v1.3) with
|
| `python verify_data.py DATA_DIR` from the repository. The checker `apidrift-check`, the lookup tool `apidrift-lookup`,
|
| the MCP server `apidrift-mcp`, the evaluation script, the integrity checks and all audit material are in the GitHub
|
| repository: <https://github.com/pushkardg/pypi-api-drift>.
|
|
|
| **Author:** Pushkar Devanahalli Gopalakrishna (Independent Researcher, USA). **Citation:** see `CITATION.cff` or the
|
| the paper, <https://doi.org/10.5281/zenodo.23648339>.
|
| |
