Kumatigi

Version v0.1.0 · released 2026-07-02 · fingerprint 88299270812cfd31 · 64,841 rows

Kumatigi is a quality-annotated parallel corpus for French–Bambara, released alongside the ACL Findings paper Kumatigi: Quality-Driven Data Augmentation for Low-Resource Machine Translation (Cissé, 2026).

Bambara is a Mande language spoken by 30–40 million people across Mali and neighbouring regions. Existing parallel resources are small, orthographically inconsistent, and mostly unlabelled for quality. Kumatigi addresses these gaps with:

• systematic orthographic normalization (special characters ɛ/ɔ/ɲ);
• per-pair quality scores enabling principled filtering;
• a dual-dataset generation framework combining round-trip synthetic pairs (fluency signal) and back-translated pairs (vocabulary coverage);
• an extended and reviewed test set with alternative Bambara references for pairs where multiple valid translations exist.

This release corresponds to the paper's Tier 1 (publicly releasable) data. Tier 2 sources (NLLB, web-mined news, private) are used to train the paper's models but are not redistributed here — please obtain them from the original providers.

Release notes for v0.1.0. Initial public release accompanying the ACL Findings 2026 paper. Human-reviewed pairs only; non-reviewed will be added once reviewed.

Splits

Config	Language pair	Train	Validation	Test	Total
bm-fr	bam↔fra	55,412	4,541	4,888	64,841

• train — releasable subset of the paper's training pool.
• validation — held-out from the corrected Bayelemabaga corpus.
• test — the Kumatigi extended test set: corrected Bayelemabaga test + alternative Bambara translations for cases with multiple valid equivalents.

Data schema

Field	Type	Description
pair_id	string	Stable 16-char SHA1 id (src_lang, tgt_lang, src_text, tgt_text).
src_lang / tgt_lang	string	ISO codes: fra, bam.
src_text / tgt_text	string	Parallel sentences.
quality_bin	class label	Quantized quality score: low [0, 0.70) · medium [0.70, 0.80) · high [0.80, 0.90) · very high [0.90, 1.0].
tier	string	Provenance tier (see below).

Tier taxonomy (tier column)

The tier encodes which side is authentic in each row, not the folder it came from. Only reviewed rows are published in this release; non-reviewed pairs are held back until annotators clear them, and future releases will grow as annotation progresses.

Tier	Meaning
high_quality	Both sides human-authored / reviewed (corrected Bayelemabaga, hand-curated pairs).
real_bam_reviewed	Bambara side is authentic; French side was model-translated and then human-reviewed.
real_fra_reviewed	French side is authentic (e.g. Project Gutenberg); Bambara side was model-translated and then human-reviewed.
synthetic_reviewed	Both sides model-generated (round-trip or back-translation) but the pair was human-reviewed and passed.

Non-reviewed variants (real_bam, real_fra, synthetic) exist in the paper's internal training pool but are not shipped here. They will be added to future releases once reviewed.

Tier distribution in train (all configs combined):

Tier	Count	%
high_quality	40,565	73.2%
real_fra_reviewed	9,943	17.9%
real_bam_reviewed	3,595	6.5%
synthetic_reviewed	1,309	2.4%

Quality-bin distribution in train:

Quality bin	Count	%
low	36,525	65.9%
medium	332	0.6%
high	1,802	3.3%
very high	16,753	30.2%

Quality bins

Scores underlying the quality_bin label are per-language-pair and per-item. For human-reviewed content they start at 0.7 by default and are elevated by native-speaker reviewers up to 1.0. For model-translated content they come from Monte Carlo dropout confidence estimation over the paper's language-pair-adapted NLLB-200 model (see paper §3.2.4). Only the quantized bin is published; the raw score is used internally for training-time weighting.

For a cleanest subset:

from datasets import load_dataset
ds = load_dataset("Tysby101/kumatigi", "bm-fr", split="train")
ds = ds.filter(lambda x: x["quality_bin"] in ("high", "very high"))

Usage

from datasets import load_dataset

# Always-latest
fr_bm = load_dataset("Tysby101/kumatigi", "bm-fr")

# Pin to this exact release
fr_bm = load_dataset("Tysby101/kumatigi", "bm-fr", revision="v0.1.0")

Sources included in this release

• Corrected Bayelemabaga — Tapo et al. 2025, released under CC-BY-SA. We fixed systematic special-character errors and refined ~2k pairs.
• Project Gutenberg → Bambara — French literary text (public domain) machine-translated with a domain-adapted NLLB-200-600M; a subset was manually reviewed.
• Text from Bambara Reference Corpus — Bambara periodicals (1970–2020) digitized via Mistral OCR + manual post-processing; French side machine-translated and reviewed.
• Contemporary community content (~1.2k pairs) — daily-observation curation of culturally significant expressions and contemporary vocabulary.
• Dictionary-derived examples (~1.6k pairs) — extracted from a French–Bambara dictionary with contextually appropriate example sentences.
• Historical treaties (Charte du Manden, Accord d'Alger) — public-domain political / historical texts.

Sources not included (Tier 2 — training only)

• NLLB dataset French–Bambara subset — Meta's license permits research use but not redistribution. Fetch directly from Meta AI.
• Web-mined news — copyright unclear; used internally for model training only.
• LLM-generated pairs pending re-review (Claude, GPT, Mistral) — held out until output-license status is clarified.
• Private items — a small pool marked private at curation time.

Ethical considerations & limitations

• Orthographic variability. Bambara has multiple scripts (Latin, Adjami, N'Ko) and inconsistent handling of special characters ɛ/ɔ/ɲ. This release standardizes on the Latin script with special characters preserved. Legacy ASCII approximations (E, O, ny) may still appear in noisier tiers.
• Model-assisted pre-annotation. Every published pair has been human-reviewed, but the initial draft came from different pipelines depending on tier: high_quality and real_*_reviewed rows have at least one authentic side; synthetic_reviewed rows were both drafted by a model (round-trip or back-translation) and then approved by a reviewer. Rows that failed review, or have not yet been reviewed, are not included in this release and will ship in a future version once annotation progresses.
• No PII scrubbing was performed beyond what was inherited from the original sources. If you find any personal data please open an issue on the Hub.

Versioning

Each release is tagged on the Hub. Pin a specific version with revision="v0.1.0" on load_dataset. The dataset_fingerprint in the front-matter is a stable SHA256[:16] over sorted pair_ids across all configs and splits — comparing fingerprints across versions detects silent drift.

A machine-readable log lives at release_history.json.

Release history

Version	Date	Total rows	Fingerprint	Notes
v0.1.0	2026-07-02	64,841	88299270812cfd31	Initial public release accompanying the ACL Findings 2026 paper. Human-reviewed pairs only; non-reviewed will be added once reviewed.

License

Released under CC-BY-SA 4.0, consistent with the original Bayelemabaga license. If you build derivatives, please retain the attribution chain (Bayelemabaga → Kumatigi → your work) and share alike.

Citation

@inproceedings{cisse2026kumatigi,
  title     = {Kumatigi: Quality-Driven Data Augmentation for Low-Resource Machine Translation},
  author    = {Ciss{\'e}, Cheick Tidiani},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year      = {2026}
}

Please also cite the foundational Bayelemabaga corpus:

@article{tapo2025bayelemabaga,
  title   = {Bayelemabaga: Creating Resources for Bambara NLP},
  author  = {Tapo, Allahsera Auguste and others},
  year    = {2025}
}

Acknowledgements

Compute for the paper's model training was provided by Orange Research. Thanks to the Bayelemabaga authors, the Corpora Mandeica initiative, and the native-speaker reviewers who assessed thousands of pairs.