# Common Voice Polish text v26.0

Text-only projection of Polish sentences associated with validated recordings
in Mozilla Common Voice Scripted Speech 26.0.

## Source and scope

- Official upstream: https://mozilladatacollective.com/datasets/cmqinmu2a00winq07gyrtri0q
- Upstream dataset ID: `cmqinmu2a00winq07gyrtri0q`
- Upstream release: `26.0` (2026-06-12)
- Pinned access mirror: `Peacockery/common-voice-scripted-speech-26@b4d8b94d43831475de59a455345acf6945cfd66e`
- Selected split: `validated`
- Language: Polish (`pl`)
- License: `CC0-1.0`
- Records: 45,043
- Tokens: 994,922 (`cl100k_base` proxy)

Audio, `client_id`, age, gender, accent, file paths and other direct contributor
metadata are not selected. Repeated recordings of the same normalized sentence
are represented once; aggregate clip count and duration are retained only in the
attribution sidecar.

## Processing

1. Read only selected metadata columns from the pinned Parquet with DuckDB.
2. Require the expected Polish locale, upstream dataset ID and CC0 license.
3. Normalize Unicode and whitespace.
4. Exclude malformed text and simple URL, email, telephone and PESEL-like matches.
5. Collapse case/punctuation-equivalent text within this source.
6. Compute deterministic IDs and `cl100k_base` token counts.

## Unit note — document length vs the corpus gate

Every record is a **single validated speech prompt** (the Common Voice unit):
45,043 docs, min 12 / median 56 / mean 58.5 / **max 153 characters** — **zero
documents reach `build_dynaword.py`'s 200-character minimum**, and nothing in
the pipeline concatenates sentences. The shard ships as
`data/<key>/<key>.parquet` (like the other in-repo sources), so the
SpeakLeash-flow gate does not run over it — but as corpus policy this is a unit
mismatch and needs a maintainer decision before merge:

- **(a) accept sentence-level units** for this source via a documented
  per-source gate override (a small shared-build change, to be reviewed as its
  own PR) plus a training-mix cap — the sentences are clean contemporary Polish
  prompts (995k tokens, ~0.01% of the corpus) and synthetic concatenation would
  fabricate documents that never existed;
- **(b) withdraw the source** as unit-incompatible with the document-level gates.

Recommendation: (a) with the cap; (b) is equally defensible given the reviewer's
lean ("single sentences may not belong here at all"). The decision is the
maintainer's.

## License at the origin and downstream terms

- License statement at the origin: the dataset record on the **Mozilla Data
  Collective** portal — <https://mozilladatacollective.com/datasets/cmqinmu2a00winq07gyrtri0q>
  — carries `license: "Creative Commons Zero v1.0 Universal (CC0-1.0)"` for the
  Common Voice Scripted Speech 26.0 datasets, and the same `license` /
  `license_url` values ride along per row in the source Parquet (the build
  rejects any row not labelled `CC0-1.0`).
- **Re-identification prohibition carried to downstream users**: the Mozilla
  Data Collective Terms of Use
  (<https://mozilladatacollective.com/terms>) forbid users to "decompile,
  analyze, or otherwise attempt to determine, trace, match the identity of, or
  re-identify the individual contributor or source of any data contained in any
  Dataset" (Data Consumer License, clause viii). Any redistribution of this
  shard must keep that restriction on its users; this projection already drops
  audio, `client_id`, age, gender, accent and file paths.

## PII scrubber and residual scan

Scrubber that ran: the `quality_rejection()` pattern filters in
`src/fetch_common_voice_pl.py` (malformed text; URL, e-mail, telephone-like and
PESEL-like matches are **rejected**, not rewritten) plus Unicode/whitespace
normalization. This is **not**
[SlayerLab/NERGAL](https://huggingface.co/SlayerLab/NERGAL); the shard must be
re-run through NERGAL before it is admitted to a release. Merge with
`release: None` is fine meanwhile.

Residual scan covers **all 45,043 records** (the shard is tiny): 0 e-mail-like
strings, 0 digit runs (7-15), 0 PESEL mentions, 0 URLs; 12 records mention the
word "telefon" in narrative prose (e.g. "w tej chwili zadzwonił telefon") — not
contact data. Residual risk is low here; the NERGAL gate stands because the
prompts are user-contributed speech material.

This source has not yet been cross-deduplicated against every released DynaWord
source. The target integration should run the corpus-wide exact and near-duplicate
protocol before a training release. Short isolated sentences should also receive
a source cap in training-mix experiments.