Commit History
6 days ago
-
Update Croissant metadata (v2 splits, RAI, NeurIPS citation)
Verified signaturebb5a567
-
Add v2 configs and documentation for already speaker-disjoint languages
Akan, Amharic, Oromo, Sidama, Tigrinya, and Wolaytta were already speaker-disjoint in their original (v1) datasets with 0% speaker leakage in eval splits. This update: 1. Adds dedicated _v2 configs in the dataset metadata for these six languages (aka_asr_v2, amh_asr_v2, orm_asr_v2, sid_asr_v2, tir_asr_v2, wal_asr_v2) pointing directly to the original data files. 2. Documents in the dataset card that these languages do not require the v2 split-map re-filtering and can be loaded directly from the original splits.5f4d8ca
-
Document the ASR v2 splits in the main dataset card
Verified signatureMoves the v2 documentation into this card, where readers will find it, and drops the separate data/ASR_v2/README.md, which nothing rendered. Two insertions only: a Table of Contents entry and a section at the end of Data Splits. The front matter and every other line are byte-identical. Condensed to corpus-level figures and the three things that actually trip people up when reading v2; per-language detail stays in data/ASR_v2/metadata/split_comparison.csv.e91442a -
Correct seven errors in the v2 card
Verified signatureAn independent verification pass over the published artefacts found the previous revision of this card wrong in seven places. * The transcript-overlap row compared two different languages as though it were one language's before and after. dga went 19.1% -> 0.2%; the 1.5% figure is ach, which is a regression from 0.6%. Both are now shown explicitly, including the regression. * The warning against reading only one v1 split claimed it reintroduces speaker leakage. It does not -- v2 moves whole speakers, so any subset of one split stays disjoint from the others. The real damage is losing 88% of the evaluation set, silently. * Streaming was described as fetching almost nothing. There is no predicate pushdown and audio is inline, so it transfers the whole labelled half: measured, Shona's 1,749-row test set costs 552 MB and ~8 minutes. * The unlabeled split was given as 816 GB in one place and 637 GB in another. It is 637 GB. * sample_rate and channels were described as mea…33e8c8d -
Document how to read the v2 splits, with verified characteristics
Verified signatureAdds a 'Reading v2 today' section with a tested recipe, and a 'What v2 guarantees' table whose every figure was measured from the published split maps joined to v1 rather than asserted from the code that produced them. Both silent-failure modes are documented and quantified: reading only the same-named v1 split returns 60 of Acholi's 515 test rows with 100% of those speakers present in v1 train, and reading `unlabeled` yields nothing after a very long wait. Also corrects three things the card got wrong: it referenced three files that were never published (the parquet metadata is blocked by the storage quota), it presented the repacked-audio schema as if it were what you read today, and it did not warn that speaker_id is reused across languages (354 collisions). excluded_rows.csv is added so the two unmapped utterances are documented.c293462 -
remove probe
Verified signaturefafb264 -
probe csv
Verified signatured9c5cac
-
Remove superseded pre-reshard copies for dag/dga/kpo/pcm
Removes the {lang}-* files these four configs pointed at before the repoint, plus the duplicate top-level dag_asr/ tree. No config references any of them; the 2026-03 reshards now serving these configs were verified row-for-row and id-for-id identical, with per-row audio payload ratios of 1.000-1.002, across every split including unlabeled. data/TTS/wol and data/TTS/yor are deliberately untouched.21bb578 -
Point dag/dga/kpo/pcm configs at the 2026-03 reshards
These four configs were still resolving to the pre-reshard filenames. The 2026-03 'Fix schema and limit shard size to 200MB' files have been in the repo since then but no glob referenced them, so users were served the older layout -- and pcm_tts's kept copy has an over-large row group that made the dataset viewer fail. Verified before the change, per split including unlabeled: identical row counts and identical id multisets, with per-row audio payload ratios of 1.000-1.002. Same utterances, same audio, better sharding. Only 15 path: lines change. All 75 comment lines, the three commented-out configs (bam_tts, fuf_tts, wol_tts) and the other 30 configs are byte-identical.3598038 -
Add ASR v2 speaker-disjoint split definitions
Adds data/ASR_v2/: the complete v2 split assignment for all 19 ASR languages, sharded per language, with the v1-vs-v2 audit tables and full provenance. Purely additive. No existing file is modified and no config is changed, so every current load_dataset call resolves exactly as it did before. Eval-set speaker leakage falls from 97-100% (in 11 of 19 languages) to 0% in every language. The split map reproduces v2 against any copy of v1 without downloading anything further.96ced60 -
Add ASR v2 split summary (probe)
2e3955e
-
Fix Kikuyu (kik) TTS scripts: update text with corrected scripts
28e04c5
-
Create croissant.json
4ab110c
-
Fix schema and manual shard for yor_tts/test
519c4b9 -
Fix schema and manual shard for yor_tts/validation
cd0836e -
Fix schema and manual shard for yor_tts/train
95e18a7
-
Fix parquet previews for ewe_asr (batch upload)
13aeb51 -
Fix parquet previews for ewe_asr (batch upload)
ff05002 -
Fix parquet previews for ewe_asr (batch upload)
62731b9 -
Fix parquet previews for ewe_asr (batch upload)
02e3b34 -
Fix parquet previews for ewe_asr (batch upload)
0fe08b9 -
Fix parquet previews for ewe_asr (batch upload)
2186755 -
Fix parquet previews for ewe_asr (batch upload)
01161e8 -
Fix parquet previews for ewe_asr (batch upload)
0b98d5b -
Fix parquet previews for ewe_asr (batch upload)
804b43e -
Fix parquet previews for ewe_asr (batch upload)
b86273c -
Fix parquet previews for ewe_asr (batch upload)
81bb8d6 -
Delete empty transcriptions for Ewe - batch 1/1
e335c80
-
Upload data/TTS/yor/yor-test-00000.parquet with huggingface_hub
8e11254 -
Upload data/TTS/yor/yor-validation-00000.parquet with huggingface_hub
1e8a182 -
Upload data/TTS/yor/yor-train-00000.parquet with huggingface_hub
a5a5050
-
Temporarily hiding Wolof, Bambara and Fula
Correcting data quality issues before re-upload8dba400
-
Update README.md
dd259e6
-
Add AIMS dataset (sharded, nested schema)
34e2fdd -
Add AIMS dataset (sharded, nested schema)
fdee46d -
Fix parquet previews for ful_tts (batch upload)
9234ef2 -
Fix parquet previews for ful_tts (batch upload)
22b7f8a -
Fix parquet previews for ful_tts (batch upload)
f5c7b43 -
Fix parquet previews for fuf_tts (batch upload)
67db8e4 -
Add fuf_tts configs
6ea1581 -
Add AIMS dataset (sharded, nested schema)
0f9e2e3 -
Remove incorrectly labeled Pular (fuf) data from 'ful' path
639a8af -
Upload README.md with huggingface_hub
b80795a
-
Fix schema and limit shard size to 200MB for wol_tts/test
21b129e -
Fix schema and limit shard size to 200MB for wol_tts/validation
d65ed41 -
Fix schema and limit shard size to 200MB for wol_tts/train
1ea9a99 -
Fix schema and limit shard size to 200MB for pcm_tts/test
b2bf610 -
Fix schema and limit shard size to 200MB for pcm_tts/validation
293edc1 -
Fix schema and limit shard size to 200MB for pcm_tts/train
a15be94 -
Fix schema and limit shard size to 200MB for kpo_asr/unlabeled
c1dc505