Add speaker-disjoint ASR re-split (v2): ach
What this is
A speaker-disjoint re-split of the labelled ASR splits of this dataset,
added under data/ASR_v2/ as new {lang}_asr_v2 configs.
Nothing existing is modified. No file outside data/ASR_v2/ is changed
except the root README.md, and that edit only appends the new config
entries -- every one of the existing configs and the entire prose body are
byte-identical, verified programmatically before this commit was built. Anyone
currently loading {lang}_asr (hackathons included) is unaffected.
The audio payloads are copied byte-for-byte from the current revision. This is
a re-partitioning, not a re-encode and not a quality filter.
Why
The v1 ASR splits share speakers between train and eval, in several languages
almost completely:
| Language | v1 eval speakers also in train | v2 |
|---|---|---|
Acholi (ach) |
99 % | 0 % |
Akan (aka) |
100 % | 0 % |
Dagbani (dag) |
98 % | 0 % |
Dagaare (dga) |
100 % | 0 % |
Ewe (ewe) |
100 % | 0 % |
Fula (ful) |
99 % | 0 % |
Ikposo (kpo) |
99 % | 0 % |
Lingala (lin) |
99 % | 0 % |
Luganda (lug) |
100 % | 0 % |
Masaaba (mas) |
98 % | 0 % |
Malagasy (mlg) |
98 % | 0 % |
Runyankole (nyn) |
98 % | 0 % |
Shona (sna) |
97 % | 0 % |
Soga (sog) |
98 % | 0 % |
A model that has heard the eval speakers in training reports a score that is
partly speaker memorisation, and the size of that effect cannot be recovered
after the fact.
Method
Speakers are the unit of assignment, so disjointness is structural rather than
enforced. Subject to that, the assignment minimises train/eval rare-content
overlap (idf-weighted, so shared common vocabulary -- which is coverage, not
leakage -- is deliberately invisible to the objective) and matches the gender
distribution of each eval split to the corpus. Deterministic from one seed;
full provenance in data/ASR_v2/metadata/splits_manifest.json.
Storage note
Hugging Face PRs push to refs/pr/N on this repository, so these LFS objects
occupy your storage from now, before you have agreed to take them. If you would
rather not host them, closing the PR and deleting the ref reclaims the space
immediately -- and data/ASR_v2/metadata/split_map.parquet (a few MB) is
enough for anyone to reproduce the split locally without the audio at all.
Happy to adjust anything, or to reduce this to the metadata-only form.
Withdrawing this for now — apologies for the noise.
Pushing the repacked audio here consumed organisation storage quota, which was my mistake: HuggingFace pull requests push LFS objects to the target repository rather than to a fork, and I did not check the storage ceiling before sending. I am deleting the ref to return the ~17.7 GB.
Nothing in the existing dataset was modified by this PR — it only added files under data/ASR_v2/, and the v1 file tree on the ref was byte-identical to main.
The substance of the work (speaker-disjoint splits for all 19 ASR languages, reproducible from a small split map) will be proposed in a form that costs essentially no storage.
refs/pr/26 ref