Datasets:
Add CC0 Polish Common Voice validated text (v26.0)
Add CC0 Polish Common Voice validated text (v26.0)
Scope
This contribution adds a text-only projection of Polish sentences associated with validated recordings in Mozilla Common Voice Scripted Speech 26.0. It does not include audio, contributor identifiers, demographic fields or file paths.
Data
- Source rows inspected:
138,798validated clip records - Unique normalized sentences retained:
45,043 - Characters:
2,637,100 - Whitespace-delimited words:
379,819 - Tokens:
994,922(cl100k_baseproxy) - License:
CC0-1.0 - Deterministic data sample: common_voice_pl.sample.jsonl
- Published source snapshot: PiotrSty/common-voice-pl-text v1.0.0
The build rejected 2,557 source rows (2,538 too short and 19 too long) and collapsed 91,198 repeated clip rows after within-source normalization.
Unit note — the 200-char gate (review item 2)
Every record is a single validated speech prompt (the Common Voice unit): min 12 / median 56 / mean 58.5 / max 153 characters — measured across all 45,043 docs, zero reach build_dynaword.py's 200-character minimum, and nothing concatenates them (no packing step exists in src/fetch_common_voice_pl.py). The shard ships as data/<key>/<key>.parquet like the other in-repo sources, so the SpeakLeash-flow gate never runs over it — but as corpus policy this is exactly the unit mismatch you flagged. Decision needed before merge:
- (a) accept sentence-level units for this source via a documented per-source gate override (small shared-build change, to be sent as its own PR) + a training-mix cap. Rationale: the data is clean contemporary Polish prompt prose (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 ("single sentences may not belong here at all" — equally defensible).
My recommendation is (a) with the cap, but the corpus-unit policy call is yours.
Sample (verbatim, from the built shard)
id: common_voice_pl_ce48ade5fc7e4c0b1ba137476aa518885e4805ff5e0cf979d840a4d4efa45b05
source: common_voice_pl
added: 2026-09-05
created: 2026-06-12
token_count: 17
license: CC0-1.0
author: Mozilla Common Voice contributors
Mniej spraw rozstrzygnięto na zasadzie polubownej
id: common_voice_pl_4c8418ce2e454c22f8518697f200beb3a4240e3e6d81b7232e48b6c3d1f75d83
source: common_voice_pl
added: 2026-09-05
created: 2026-06-12
token_count: 26
license: CC0-1.0
author: Mozilla Common Voice contributors
— Oni używają redakcyjnego „my” — zaznaczył naraz Nowak.
id: common_voice_pl_8dd4e66c55fada3f8253a5e061f460325e5d66ad5c763d0fda9a3744fa02492c
source: common_voice_pl
added: 2026-09-05
created: 2026-06-12
token_count: 14
license: CC0-1.0
author: Mozilla Common Voice contributors
W tym okresie powstały również nierównowagi
(Yes — these are the documents; the sample doubles as the length-gate exhibit.)
License at the origin (review item 3)
The CC0-1.0 grant is stated in the Mozilla Data Collective dataset record at the origin: https://mozilladatacollective.com/datasets/cmqinmu2a00winq07gyrtri0q — its license field reads 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 — unexpected_license contract check). The pinned mirror (Peacockery/common-voice-scripted-speech-26@b4d8b94d…) is only an access path to that record's data.
Terms — re-identification prohibition (review item 4)
The Mozilla Data Collective Terms of Use (https://mozilladatacollective.com/terms, Data Consumer License) 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" (clause viii). This is carried to downstream users: the datasheet and NOTICE state the prohibition as a redistribution condition, and the projection itself drops audio, client_id, age, gender, accent and file paths.
PII (review item 6)
Scrubber that ran: the quality_rejection() pattern filters in src/fetch_common_voice_pl.py — malformed text and URL / e-mail / telephone-like / PESEL-like matches are rejected, not rewritten — plus Unicode/whitespace normalization. This is not SlayerLab/NERGAL. Residual scan over all 45,043 records: 0 e-mail-like, 0 digit runs (7–15), 0 PESEL mentions, 0 URLs; 12 records say "telefon" in narrative prose ("w tej chwili zadzwonił telefon") — not contact data. No release before a NERGAL pass (the prompts are user-contributed speech material); merge with release: None is fine meanwhile. Stated in the datasheet.
Provenance (original)
- Official dataset:
Mozilla Data Collective / Common Voice Scripted Speech 26.0 - Polish - Official dataset ID:
cmqinmu2a00winq07gyrtri0q - Release date:
2026-06-12 - Pinned access mirror:
Peacockery/common-voice-scripted-speech-26@b4d8b94d43831475de59a455345acf6945cfd66e - Source Parquet LFS SHA-256:
abd846a0411d04cf88920161100625e9cbe64a111d17474ac49375f490dafb3a - Immutable prepared payload:
PiotrSty/common-voice-pl-text@2e3e62d374df8cad50e443a356e2c2bec86e10b4 - Prepared Parquet SHA-256:
2c22fb1be8badb75f55c0f1f6719360ab50dc798054a7673d67b2780c41858b2
The source is read with DuckDB column projection, so the 4.48 GB embedded audio payload is never downloaded or copied into this contribution.
Quality and privacy (original)
- Unicode NFKC and whitespace normalization
- Casefolded punctuation-insensitive exact deduplication within the source
- Simple malformed-text, URL, email, telephone-like and PESEL-like filters
- Deterministic
langid.pyaudit - Audio,
client_id, age, gender, accent and file paths not selected; repeated recordings collapse to one sentence record (aggregate clip count/duration kept only in the attribution sidecar)
Cross-source exact/near dedup and benchmark checks remain pending; short isolated sentences should get a training-mix cap in experiments.
Registry (review item 1)
Rebased onto main @ 34e7ea3; the src/sources.py diff is exactly the registry entry — +23 / −0, no whole-file rewrite (line-ending note in the review thread). The entry carries added: "2026-09-05", release: None, the origin license link and the re-identification covenant.
Test plan
-
python3 -m pytest src/test_common_voice_pl_contract.py— passed on the rebased branch - Length-gate measurement (all 45,043 docs; decision options above)
- License verified at the origin (MDC dataset record + per-row
license) - Re-identification covenant sourced to MDC Terms clause (viii) and carried downstream
- Residual PII scan over the full shard (clean)
- Reviewer/maintainer: unit decision — (a) accept sentence units + cap + separate gate-override PR, or (b) withdraw
Review against AGENTS.md ("What the PR must show", "Normalize the text before you build"). Needed before merge:
- Line endings.
src/sources.pyis committed with CRLF, so it shows as a whole-file rewrite (the real change is 17 lines) and will conflict with every other source PR. Please convert to LF. - Length gate. ~45k docs / 995k tokens is ~58 chars per doc, but
build_dynaword.pydrops docs under 200 chars. Explain how these pass (concatenation?) — single sentences may not belong here at all. - License at the origin. Link the license statement on the Mozilla Common Voice dataset page (CC0), not only your HF mirror — the mirror can't be evidence for itself.
- Terms. Common Voice terms forbid attempts to re-identify speakers — say how that's carried to downstream users.
- Sample inline. Paste 2–3 real documents into the description, one fenced block each: metadata as
key: valuelines, a blank line, then the text with its real line breaks, truncated with[…](format as in #24). A linked.sample.jsonlrenders on the Hub as one escaped line per document and can't be reviewed at a glance. - PII. AGENTS.md requires SlayerLab/NERGAL. The source can merge with
release: Noneafter another scrubber, but it won't be admitted to a release until the shard is re-run through NERGAL. State in the datasheet which scrubber ran.
Review addressed — pushed as a rebase + review-fix commit on refs/pr/22, onto main @ 34e7ea3. Point by point:
- Line endings. The
src/sources.pydiff is now exactly the registry entry: +23 / −0 — no whole-file rewrite. (Measurement note:main's own blob is CRLF today —git cat-file blob origin/main:src/sources.py= 533/533 CR; matchingmain's bytes is what minimizes the diff, converting this file to LF against a CRLFmainwould recreate the rewrite. Repo-wide LF normalization is maintainer work — say the word and I'll send it as its own PR.) The entry now also carriesadded: "2026-09-05"andrelease: None. - Length gate — the honest answer: they don't pass, and nothing concatenates them. Measured over all 45,043 docs: min 12 / median 56 / mean 58.5 / max 153 chars — 0 docs reach
build_dynaword.py's 200-char minimum;src/fetch_common_voice_pl.pyhas no packing/concatenation step (the unit is the single validated speech prompt — the Common Voice unit itself). The shard ships asdata/<key>/<key>.parquetlike the other in-repo sources, so the SpeakLeash-flow gate never touches it — but as corpus policy this is exactly the unit mismatch you suspected. Decision needed: (a) accept sentence-level units via a documented per-source gate override (small shared-build change → its own PR) + training-mix cap — rationale: clean contemporary Polish prompt prose, 995k tokens ≈ 0.01% of the corpus, and synthetic concatenation would fabricate documents that never existed; or (b) withdraw as unit-incompatible ("single sentences may not belong here at all" — equally defensible). My recommendation is (a)+cap; the corpus-unit call is the maintainer's. The sample below doubles as the exhibit. - License at the origin. The CC0-1.0 grant is stated in the Mozilla Data Collective dataset record: https://mozilladatacollective.com/datasets/cmqinmu2a00winq07gyrtri0q —
license: "Creative Commons Zero v1.0 Universal (CC0-1.0)"— and thelicense/license_urlvalues ride along per row in the source Parquet (the build rejects any non-CC0 row —unexpected_licensecontract check). The pinned mirror is only an access path. - Terms — re-identification. Sourced to the MDC Terms of Use (https://mozilladatacollective.com/terms, Data Consumer License clause viii): users may not "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". Carried to downstream users in the datasheet + registry
traceableas a redistribution condition; the projection itself drops audio,client_id, age, gender, accent and file paths. - Sample inline. 3 real documents in the description, one fenced block each (
key: valuemetadata, blank line, real line breaks, format as in #24) — they are single sentences, which is also the length-gate exhibit. - PII. Datasheet names the scrubber: the
quality_rejection()pattern filters insrc/fetch_common_voice_pl.py(malformed text and URL/e-mail/telephone-like/PESEL-like matches are rejected) + normalization — not NERGAL. Residual scan over all 45,043 records (the shard is tiny): 0 e-mail-like, 0 digit runs, 0 PESEL mentions, 0 URLs; 12 narrative "telefon" mentions ("w tej chwili zadzwonił telefon"), not contact data. Agreed: no release before a NERGAL pass (user-contributed speech prompts); merge withrelease: Noneis fine meanwhile.
Verification: python3 -m pytest src/test_common_voice_pl_contract.py — passed on the rebased branch.