PiotrSty/yodas3-transcripts-filipino

YODAS Filipino transcripts (yodas3_fil) - spoken-language captions

Filipino caption tracks from YODAS v3, one complete transcript per record. This text-only snapshot covers every Filipino metadata shard at the pinned source revision. English translations and word-level duplicate annotations are excluded from the text.

• Source records: 6,790 across 16 shards
• Retained records: 6,565 (96.7%)
• Measured cl100k_base proxy tokens: 21,565,408
• Characters: 63,775,326
• Before exact deduplication: 21,607,107 tokens
• Added: 2026-10-08; original creation dates are unknown
• Source: espnet/yodas3
• Source revision: 81d802f45a1c1e1f5331260484dacf1fbea81981
• Status: public research candidate, original attribution under review

License - evidence at the origin

The publisher's pinned dataset card declares CC BY 3.0 for the source recordings and dataset. The per-record license field preserves this declaration.

Original author, title and film URL are absent from the source metadata. These fields are not reconstructed or attributed to the dataset curators. artifacts/attribution.jsonl.gz records known provenance, hashes and the missing attribution fields for every retained record. The owner authorized public access on 2026-10-08. Original attribution and training clearance remain unresolved; release remains null. See artifacts/publication_event.json for this access change.

Samples

Download 12 complete sample records. Samples are selected deterministically by SHA256 of record IDs across the full retained corpus. The three excerpts below use those exact records; [...] marks truncation. These are source captions and may contain transcription errors.

Sample 1

id: yodas3_fil:9b2df4190ae1a0a9113c1c100edb8edc
source: yodas3_fil
created: null
token_count: 2041

[Musika]
sang una ko ikaw
nakita ako
nagtalang nakasulod sa isah
daluhan
ikaw
nango
kamatay
sa apang
karon ina lang ang masiling
mo
sorry
[Musika]
sorry I am so
sorry Okay just go
apang h
mag hindi mo ako makita
nagang
[Musika]
ak
[Musika]
saod
angb ka sa
[Musika]
gamp mo
[Musika]
ako
sorry I love
[Musika]
you hal
ako
isang TP
mga tinaga ng masakit
mas masad halin sa mga lamharon
kasingkasing nagahigugma
tanan Mak mga pin isa
[Musika]
Diary de Di [...]

Sample 2

id: yodas3_fil:f821b846b9470ba2b217b0cc734bdf7a
source: yodas3_fil
created: null
token_count: 1170

Ang nangiti ni Aling Pasing. Tahimik
lang ang barangay namin sa Tagaytay. Ang
pinakamaraming ingay dito ay ang simoy
ng hangin at huni ng mga ibon. Dito ako
nagretiro sa maliit na bahay na puno ng
mga halaman. Ang akala ko, puro
kapayapaan na lang ang haharapin ko.
Pero nagkamali ako. Dumating si Aling
Pasing, "Mga anim na buwan na ang
nakakaraan." Bumili siya ng lumang bahay
sa tapat lang ng kalsada. Matanda na rin
siya. Parang nasa pit na. Pala [...]

Sample 3

id: yodas3_fil:f342c1f24f955d4160a40318332abfe4
source: yodas3_fil
created: null
token_count: 2680

[musika]
[musika]
una ko ikaw nakita
naglaom ako nagtalang [musika] ako kag
nakasulod sa isa ka matah ng dalamkuhan
[musika]
ikaw ang babae madugay ko ginapangita ka
ginahul
nagpangako [musika] ka
nga kamatayo lang ang makapabulag sao
kag sa akang
karon [musika]
ina lang ang masiling mo
sorry
[musika]
>> ah
>> nga ah
[musika]
>> i am so Sorry.
>> Okay, just go
>> apang hindi na magbalik.
[musika]
Hindi mo ako makita ng nagahibi.
Bangod ak mga luh [...]

Filtering

All source shards pass SHA256 verification against the pinned upstream manifest. Caption segments retain their original order; leading/trailing segment whitespace is trimmed and segments are joined with newlines. Empty tracks are excluded. Exact deduplication uses the full SHA256 of each complete transcript and keeps the first occurrence in source order.

Final exclusions, one decision per source record in artifacts/decisions.jsonl.gz:

{
  "empty": 1,
  "exact_duplicate": 224
}

The fil language label comes from the original caption track. Independent language detection, PII review, near-deduplication and benchmark overlap checks have not been performed. Their QA values are marked as unmeasured.

Record fields

The primary columns follow the same layout as the other PiotrSty corpora:

Field	Meaning
id	Stable transcript identifier
text	Full joined caption track
source	yodas3_fil
added	Date of the initial local export, 2026-10-08
created	Null: original recording/upload date is unknown
token_count	Measured cl100k_base tokens
license	CC-BY-3.0, as declared by the publisher
author	Null: original author is not supplied
source_url	Immutable HF source-shard link
title	Null: original film title is not supplied

Additional columns retain language, text hash, source record ID, repo revision, source path, duration, locale and review status. source_url is not the original film URL. The train split follows the account's loading convention and does not imply training approval. The existing publication_approved=false field records the pipeline's unresolved review status, not the repository's current visibility.

Files

• data/train-00000-of-00001.parquet: complete corpus.
• artifacts/sample.jsonl: 12 complete records.
• artifacts/attribution.jsonl.gz: provenance and attribution gaps.
• artifacts/decisions.jsonl.gz: 6,790 retention/exclusion decisions.
• artifacts/stats.json, qa.json, run.json, ontology.json: research artifacts.
• artifacts/checksums.json: SHA256 manifest.
• src/: acquisition, export, packaging code, tests and card template.

Reproduction

Install src/tagalog_literature_requirements.txt. In a workspace, place the contents of src/ under scripts/. Restore artifacts/source_manifest.json as data/yodas3_fil_size_audit_v1/manifest.json and artifacts/upstream_README.md as data/yodas3_fil_size_audit_v1/source_README.md to pin acquisition to this snapshot. Then run from the workspace root:

python scripts/audit_yodas_fil_size.py
python scripts/export_yodas_fil.py
python scripts/package_publish_yodas_fil.py

The acquisition script downloads metadata and transcriptions only. Verification reconstructs every retained text, recalculates tokens and hashes, checks complete decision coverage, and matches samples to the Parquet records. The packaging step writes locally unless --upload-private is explicitly supplied.

Known limitations

• Automatic captions can contain errors, repetitions and Taglish/other languages.
• Exact document dedup does not remove repeated segments or near duplicates.
• Original author/title/film-URL attribution is incomplete.
• PII and contamination reviews remain pending; the corpus is not anonymized.
• Token totals are tokenizer-specific measurements before additional quality filtering.

See artifacts/package_validation.json for technical verification. Public access and successful verification do not resolve the pending attribution or training decision.