Datasets:
Lingala Combined Parallel Corpus
This dataset contains 702,510 aligned Lingala–French text pairs in the lingala_french subset. The Lingala text is always in source; its French translation is in target. The data was assembled from existing OPUS packages and authored examples scraped from dic.lingala.be.
This is a structurally cleaned collection, not a human-verified translation benchmark. Some records are short expressions, interface strings, or paragraphs rather than individual sentences.
Subsets and schema
| Subset | Direction | Split | Samples |
|---|---|---|---|
lingala_french |
ln-fr | train |
702,510 |
| Total | 702,510 |
Each UTF-8 TSV has exactly three string columns and a header row:
| Column | Meaning |
|---|---|
id |
Unique identifier: ln-{target language}-{corpus slug}-{7-digit sequence}. |
source |
Lingala text. This field is text, not a corpus name or URL. |
target |
French translation. |
IDs are unique throughout the dataset. The sequence is local to each originating corpus and language pair. IDs are reproducible for unchanged input order and filtering rules; they are not permanent identifiers across future rebuilds. Row IDs were regenerated for this release.
There is one train split per subset and no validation or test split. The split name describes packaging, not a guarantee that the text is suitable for model training.
Source contributions
Counts below describe the actual rows retained in these TSV files, after filtering and deduplication. They are not the sizes advertised by the upstream datasets. Each retained pair is assigned to exactly one originating corpus; the source counts sum to the subset totals without double-counting overlaps.
lingala_french
| Originating corpus | Version / collection | ID corpus slug | Retained samples |
|---|---|---|---|
| Mozilla-I10n | v1 | mozilla |
2,349 |
| NLLB | v1 | nllb |
691,023 |
| Tatoeba | v2026-07-08 | tatoeba |
56 |
| dic.lingala.be | 2026-11-21/22 crawl | dictionary |
1,546 |
| wikimedia | v20260327 | wikimedia |
7,536 |
| Subset total | 702,510 |
Counting and overlap policy
Exact pairs are deduplicated within each language pair. When a pair occurs in multiple corpora, the first source in the collection order receives the retained-row attribution. French source order: Mozilla-I10n, NLLB, Tatoeba, dic.lingala.be, wikimedia. Distinct translations of the same Lingala text remain separate rows.
The French dictionary contribution is smaller than its 1,599 extracted unique examples because examples already present in earlier sources are counted there. Its 3,397 lexical glosses are not included. The dictionary crawl retrieved 4,817 pages by following French entry links from five seed words (mbote, kolinga, kosala, moto, mai); it is not a complete dictionary database export.
Collection and processing
Collection took place on 2026-11-21/22. OPUS Moses files were paired line by line with equal-length checks. Mozilla raw XML sentences were joined using the official XCES alignment IDs. Dictionary examples used the webpage’s explicit Lingala/French dt/dd pairs; translations were not generated by an AI system.
Processing applies NFC Unicode normalization and whitespace collapse, removes invisible Unicode format characters, and drops empty rows, encoding replacement characters, remaining control characters, rows with no letters on either side, case-insensitive identical source/target text, and exact duplicate pairs. Lingala diacritics are preserved. This build reuses the provenance-bearing combined candidate files and performs the final cleaning again.
| Subset | Input candidates | Removed in final cleaning | Retained |
|---|---|---|---|
lingala_french |
702,601 | 91 | 702,510 |
Detailed counts, removal reasons, unique text counts, whitespace-token counts, file sizes, and SHA-256 hashes are provided in statistics.json. Counts exclude header rows.
Excluded resources
- TICO-19 is retained separately in the collection as evaluation data and is not included here. Exact case-insensitive Lingala matches with that held-out data were excluded at candidate export time; fuzzy overlap and overlap with other benchmarks have not been checked.
- FLORES+ files were unavailable without authenticated access and are not included.
- QED was quarantined after samples appeared to be Turkish rather than the advertised language pair.
- Dictionary glosses and YouTube material are not included.
Licensing and attribution
No single dataset-wide license is asserted. The notices below describe the downloaded source packages; they do not relicense the underlying texts. The source-specific notices are retained under notices/. The ID corpus slug links each row to the applicable source entry.
| Source | Notice supplied with source | Preserved notices |
|---|---|---|
| Mozilla-I10n | MPL 2.0 | Notices |
| NLLB | ODC-By 1.0; original-source terms also apply | Notices |
| Tatoeba | CC BY 2.0 FR (downloaded OPUS package) | Notices |
| dic.lingala.be | No reuse license stated on retrieved pages | No license notice found |
| wikimedia | CC BY-SA 4.0 | Notices |
Publication review remains necessary for dic.lingala.be: the retrieved pages do not state a reuse license, and collection does not establish permission to redistribute its examples. Resolve that permission or exclude those examples and regenerate the counts before public distribution. NLLB also references the terms of its original text sources. The compilation has not undergone a complete rights or attribution audit.
The dataset-card license metadata is intentionally omitted because the corpus has mixed source conditions and unresolved dictionary reuse terms. See each upstream source and its preserved notices for attribution and citation details.
Quality and limitations
No additional language-identification model or semantic alignment verifier was run on this compilation. Upstream language labels can be wrong, and some sampled translations are questionable. Localization placeholders, mixed-language text, names, fragments, multi-sentence units, and noisy web text may remain. Mined data may include machine translations. Domain balance is highly uneven, especially the large NLLB contribution. There has been no comprehensive personal-data, offensive-content, or bias audit. Use the counts as structural statistics, not as an estimate of verified correct translations.
Loading
The following shows how to use this dataset:
from datasets import load_dataset
fr = load_dataset("IndabaXRDC/lin_fr_700k_v00", "lingala_french", split="train")
To load a local TSV while preserving literal strings such as NA:
from datasets import Features, Value, load_dataset
data = load_dataset(
"csv",
data_files="lingala_french/train.tsv",
sep="\t",
keep_default_na=False,
features=Features({name: Value("string") for name in ("id", "source", "target")}),
split="train",
)
The front matter declares the configuration explicitly, following the Hugging Face configuration documentation.
Acknowledgments
The present corpus has been compiled by IndabaX DRC. Credit also belongs to the original text authors, translators, dictionary author Peter Van Hertum, upstream corpus teams, and OPUS. Source READMEs under notices/ preserve their references. For OPUS, cite Jörg Tiedemann (2012), Parallel Data, Tools and Interfaces in OPUS.
- Downloads last month
- 36