dawiet/ByeByePerspectiveAPI-snapshot

ByeByePerspectiveAPI -- Rescoring Snapshot

Companion data release for "Bye Bye Perspective API: Lessons for Measurement Infrastructure in NLP, CSS and LLM Evaluation" (EMNLP 2026 Findings).

Perspective API's discontinuation at the end of 2026 leaves behind non-updatable benchmarks and irreproducible results across a large body of NLP, computational social science, and LLM-evaluation research. This repository is a snapshot archive: 77 publicly available toxicity/hate-speech datasets and major benchmarks, rescored with Perspective API before shutdown, released as a frozen, versioned, fully-documented reference point for reproducibility going forward.

What's here

• datasets/<key>/ -- one folder per dataset. Each contains <key>.parquet (the scored data) and its own README.md (source, license pointer, merge-back instructions, and any dataset-specific caveats).
• datasets_overview.csv -- one row per dataset: corresponding paper, dominant language(s), usage category, and row count. See the paper's Appendix for the full table.
• literature_review/ -- the bibliographic review this release is grounded in: which papers use Perspective API, how they were found, and how they were categorized. See literature_review/README.md.

Data schema (every dataset parquet)

column	meaning
dataset_id	this dataset's key
input_id	the original dataset's own row id (or split-local index if none existed) -- use this to merge back onto the original release
toxicity, severe_toxicity, identity_attack, insult, profanity, threat	Perspective production-attribute scores (0-1), or null if not requested/failed
lang	source-declared language code (per-row where available)
detected_languages	language(s) Perspective itself detected for the request
ok	whether the API call succeeded
error	failure reason where ok is false
query_ts	when this row was scored -- Perspective exposes no public model-version string, so this is the closest available version pin

No text is included. This release ships only identifiers, Perspective scores, and query metadata -- never the underlying text -- so redistribution stays within each original dataset's own license/terms. Perspective scores in this repository are released under CC-BY 4.0; the identifiers let you re-attach the original text yourself.

Merging scores back onto the original text

Every dataset's own README.md (under datasets/<key>/) gives its exact source and a runnable code snippet, but the pattern is the same everywhere:

1. Re-download the original dataset from the source named in that dataset's README (Hugging Face, GitHub, OSF, Zenodo, or figshare).
2. Build an input_id column on the original data -- its own id column if it had one, otherwise its 0-based row index within the split.
3. Join on input_id:

import pandas as pd

scores = pd.read_parquet("datasets/<key>/<key>.parquet")
original = ...  # load however that dataset is normally loaded
original["input_id"] = original["input_id"].astype(str)  # see that dataset's README for the exact id column
scores["input_id"] = scores["input_id"].astype(str)
merged = original.merge(scores, on="input_id", how="left")

Exact-duplicate texts within a dataset were collapsed before scoring, so a duplicate row in the original data won't have its own input_id -- look its score up by exact text match against a row that does have one.

Methodology summary

• Endpoint: https://commentanalyzer.googleapis.com/v1alpha1/comments:analyze, six production attributes (toxicity, severe toxicity, identity attack, insult, profanity, threat), doNotStore: true on every request.
• Language hints: sent only for Perspective's documented production-supported languages; otherwise the request relied on Perspective's own auto-detection.
• Query window: 2026-05-22 through 2026-08-29.
• Dataset discovery: (i) known, curated Hugging Face hate-speech/toxicity releases; (ii) a scan of every GitHub/OSF/Zenodo/figshare repository already retrieved as a code artifact of a paper in the literature review corpus, using both filename- and content-based text-column detection; (iii) a targeted search, per included paper without an identified code artifact, for its corresponding public release. Every dataset was manually checked against its cited source paper before inclusion; unverifiable matches were excluded rather than guessed.
• Coverage: 6,162,826 total rows, 5,897,618 (95.7%) scored successfully. Failures concentrate in languages Perspective does not support at all (e.g. Turkish, Finnish) -- reported explicitly per dataset rather than hidden.

See the paper's appendix and literature_review/README.md for full detail.

Citation

If you use this snapshot, please cite the paper:

@inproceedings{byebyeperspectiveapi2026,
  title     = {Bye Bye Perspective API: Lessons for Measurement Infrastructure in NLP, CSS and LLM Evaluation},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026},
  year      = {2026}
}