This repository mirrors text from third-party corpora. The upstream compilations and annotations are released by their authors under their own terms, but the underlying content is not ours to relicense: MasakhaNEWS and XL-Sum reproduce article bodies from BBC, VOA and other outlets, and AfriSenti reproduces tweet text from identifiable people. We therefore gate access rather than redistribute openly. If you only need the corpora themselves, please obtain them from the upstream sources linked in the card below instead of requesting access here.

CoSDA Source Corpora — gated mirror

The exact source-corpus snapshots used while building CoSDA, the artifact for the EMNLP 2026 Findings paper "When Audit Quality Fails to Predict Downstream Utility: Synthetic-Data Selection for Low-Resource Classification" (Le, Tran-Truong, and Ha Xuan).

Paper artifact and code: https://github.com/lexuanbach/CoSDA
152,716 rows across 5 corpora, 4 languages (Amharic, Hausa, Swahili, Yorùbá), 51 JSONL files

This is a mirror, not a new dataset. It exists so that a specific set of downloads can be reproduced byte-for-byte later, after upstream cards, configs or revisions have moved. It adds no annotation and no new text. If you want these corpora, get them upstream — the links are in the table below. Request access here only if you specifically need the snapshot this paper used.

Why this repo is gated

The CoSDA artifact deliberately does not redistribute source-corpus text, and says so in its README and DATA_LICENSE.md. The reason is upstream rights rather than the dataset licences: MasakhaNEWS and XL-Sum reproduce article bodies from BBC, VOA and other outlets, and AfriSenti reproduces tweet text. Neither the corpus authors nor we hold those rights, and X's developer terms cover tweet text separately from any CC grant. Gating this mirror keeps the snapshot recoverable for reproduction without turning it into an open redistribution of that content.

Nothing in the artifact's reproduce.sh needs this data. It is required only to re-run generation or downstream training from scratch.

Subsets

Each subset is {dataset}_{lang} with train / validation / test splits.

Config prefix	Upstream source	Languages	Rows	Fields
`masakhanews_*`	`masakhane/masakhanews`	amh, hau, swa, yor	9,469	`text`, `label`
`afrisenti_*`	`shmuhammad/AfriSenti-twitter-sentiment`	amh, hau, swa, yor	49,770	`text`, `label`
`masakhaner2_*`	`masakhane/masakhaner2`	hau, swa, yor	27,406	`tokens`, `ner_tags`
`massive_*`	`AmazonScience/massive`	amh, swa	33,042	`text`, `intent`, `annot_utt`
`xlsum_*`	`csebuetnlp/xlsum`	amh, hau, swa, yor	33,029	`text`, `summary`

masakhaner2 has no Amharic config upstream, so it covers three languages here.

Only masakhanews and afrisenti are used in the paper — they are the two benchmarks behind its 8 task-language cells. masakhaner2, massive and xlsum were pulled during scoping, never entered the reported experiments, and are included only so the snapshot is complete. Do not read their presence as a claim about the paper's coverage.

Row counts

Subset	train	validation	test
`masakhanews_amh`	1,311	188	376
`masakhanews_hau`	2,219	317	637
`masakhanews_swa`	1,658	237	476
`masakhanews_yor`	1,433	206	411
`afrisenti_amh`	5,984	1,497	1,999
`afrisenti_hau`	14,172	2,677	5,303
`afrisenti_swa`	1,810	453	748
`afrisenti_yor`	8,522	2,090	4,515
`masakhaner2_hau`	5,716	816	1,633
`masakhaner2_swa`	6,593	942	1,883
`masakhaner2_yor`	6,876	983	1,964
`massive_amh`	11,514	2,033	2,974
`massive_swa`	11,514	2,033	2,974
`xlsum_amh`	5,761	719	719
`xlsum_hau`	6,418	802	802
`xlsum_swa`	7,898	987	987
`xlsum_yor`	6,350	793	793

Loading

from datasets import load_dataset

# needs `huggingface-cli login` with an account that has been granted access
news = load_dataset("lexuanbach/cosda-source-corpora", "masakhanews_swa")
sent = load_dataset("lexuanbach/cosda-source-corpora", "afrisenti_hau", split="train")

Label fields are the upstream integer encodings, unchanged. Consult the upstream card for each subset's label mapping; this mirror does not restate them and should not be treated as authoritative for them.

Provenance

Downloaded with the HF datasets library in May 2026 and written out verbatim as JSONL, one record per line, with no filtering, deduplication, reordering, or text normalisation. Only the container changed. The download log — upstream id, config, split, row count and the one failed config — ships in the artifact repo as data/manifest/download_status.json.

Because the upstream loaders may have changed since, a fresh download today is not guaranteed to match this snapshot. That mismatch is exactly what this mirror is for.

Licensing

The per-subset terms are the upstream terms; nothing here relicenses them. The repo-level tag is cc-by-nc-sa-4.0 because that is the most restrictive of the mixed set — it is a floor for the collection, not a statement about any individual subset.

Subset	Terms
`masakhanews_*`	CC BY-NC 4.0 per the dataset card; Adelani et al. (2023) state release "under academic license or CC BY-NC 4.0". The repo's YAML tag reads `afl-3.0`, which contradicts the card prose; we follow the more restrictive reading.
`afrisenti_*`	CC BY 4.0 per the authors' own repository and dataset card. The `masakhane/afrisenti` mirror carries a conflicting `cc-by-nc-sa-2.0` tag that we believe is a metadata error; we flag it rather than silently resolve it.
`masakhaner2_`, `massive_`, `xlsum_*`	See the upstream card linked in the subsets table. We have not independently audited these three and deliberately do not restate a licence string for them here.

These grants cover the annotations and the compilation. They do not reach the underlying article bodies or tweet text, and neither do we — see "Why this repo is gated" above.

If you believe your rights are affected by this mirror, open a discussion on this repo and it will be taken down.

Citation

Cite the upstream corpora you actually use. For the snapshot and the work it supports:

@inproceedings{le2026cosda,
  title     = {When Audit Quality Fails to Predict Downstream Utility:
               Synthetic-Data Selection for Low-Resource Classification},
  author    = {Le, Xuan-Bach and Tran-Truong, Phat T. and Ha Xuan, Son},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026},
  year      = {2026}
}

Upstream: Adelani et al. 2023, MasakhaNEWS, IJCNLP-AACL (https://aclanthology.org/2023.ijcnlp-main.10/) · Muhammad et al. 2023, AfriSenti, EMNLP (https://aclanthology.org/2023.emnlp-main.862/) · Adelani et al. 2022, MasakhaNER 2.0, EMNLP (https://aclanthology.org/2022.emnlp-main.298/) · FitzGerald et al. 2023, MASSIVE, ACL (https://aclanthology.org/2023.acl-long.235/) · Hasan et al. 2021, XL-Sum, Findings of ACL (https://aclanthology.org/2021.findings-acl.413/).

Downloads last month: 11

Total file size:

158 MB