maseyan/French_Medumba

Dataset Card for French–Medumba Parallel Corpus

A small parallel corpus of French ↔ Medumba (byv) sentence pairs, intended as a seed resource for machine translation, language-learning tools, and documentation of the Medumba language of West Cameroon.

Dataset Details

Dataset Description

Medumba is a Bamileke language of the Grassfields Bantu branch, spoken primarily in the Ndé division of the West Region of Cameroon (Bangangté and surrounding villages). It is a tonal language. In this dataset it is written using the General Alphabet of Cameroonian Languages (AGLC), with tone diacritics and special characters (ɑ, ɛ, ɔ, ŋ, ʉ, ə).

The corpus contains 510 unique French–Medumba pairs, derived from 468 unique French source sentences (some French sentences have multiple acceptable Medumba translations, which are stored as separate rows). Content covers everyday vocabulary, common phrases, questions, imperatives, numerals, and a small set of grammatical paradigms illustrating Medumba tense and aspect.

• Curated by: Yannick KG
• Funded by: Yannick KG
• Shared by: Yannick KG
• Language(s) (NLP): French (fr), Medumba (byv)
• License: CC BY 4.0
• Contact: maseyan@yahoo.fr

Dataset Sources

• Repository: [Will be added once the dataset is published on the Hugging Face Hub]

Uses

Direct Use

• Fine-tuning multilingual MT models (e.g., NLLB-200, M2M-100) to add or improve Medumba coverage. The corpus is too small to train an MT system from scratch, but can noticeably shift a multilingual model's behavior on a new target language.
• Evaluation / benchmarking for fr↔byv translation quality. Using only the test split for reporting is recommended.
• Language-learning applications (flashcards, spaced repetition, phrasebooks) for French speakers learning Medumba.
• Linguistic documentation and tokenizer / orthography research for Grassfields languages.

Out-of-Scope Use

• Training a production translation system from scratch. 510 pairs is orders of magnitude below what is needed.
• Long-form or formal-register generation. The corpus contains almost no multi-clause narrative, legal, medical, or technical content.
• Claims about Medumba dialect variation. The dataset does not annotate regional dialect (Bangangté, Bazou, Tonga, etc.) and should not be used to study inter-dialectal differences.
• Automatic speech recognition / TTS without independent audio data — this is a text-only corpus.

Dataset Structure

Each row has the following fields:

• translation (dict): an object with two string keys
  • fr: the French source sentence
  • byv: the Medumba translation (AGLC orthography, NFC-normalized)
• source_id (int): 1–511, the line number in the original compiled source list, useful for tracing a pair back to its origin
• domain (str): a coarse auto-assigned tag, one of vocabulary, phrase, question, sentence, numbers, grammar_examples

Splits:

Split	Pairs
train	424
validation	40
test	46

Split integrity: splits are produced at the level of unique French source sentences (not at the level of pairs). This guarantees that when a French sentence has multiple valid Medumba translations, all of them stay in the same split. There is zero French-source overlap between train, validation, and test.

Example row:

{
  "translation": {
    "fr": "Je t'aime",
    "byv": "Mə̀ kɔ̀ o"
  },
  "source_id": 216,
  "domain": "sentence"
}

Dataset Creation

Curation Rationale

Low-resource languages of Cameroon — and Grassfields Bantu languages in particular — are severely underrepresented in NLP datasets and in publicly available MT systems. Medumba has no widely-used parallel corpus on the Hugging Face Hub. This dataset was created to provide a small but clean, schema-correct seed that other researchers, educators, and the Medumba-speaking community can build on, benchmark against, and extend.

Source Data

Data Collection and Processing

The pairs were compiled from Medumba educational material: vocabulary lists, common phrases, numerals, and small grammatical paradigms of the kind used in beginner-to-intermediate language teaching.

Processing pipeline:

1. Parse numbered N. French : Medumba lines with a regex.
2. Unicode-normalize both sides to NFC — critical for tonal orthography, because combining tone marks can otherwise be encoded in multiple equivalent forms that a tokenizer will treat as distinct tokens.
3. Collapse internal whitespace.
4. Expand variant translations separated by / (e.g.,  wə? / À bə α̂ wə?) into separate rows sharing the same French source.
5. Remove exact (fr, byv) duplicates.
6. Auto-tag with a coarse domain label based on lightweight heuristics on the French source.
7. Split into train/validation/test at the level of unique French source sentences to prevent leakage.

Who are the source data producers?

The French–Medumba pairs reflect teaching content produced by Medumba speakers and educators in Cameroon. Specific attribution: [Add here the names or sources of the materials you compiled from, if known — otherwise keep as "compiled from openly shared Medumba teaching materials"].

Annotations [optional]

Annotation process

The only automatic annotations are:

• source_id: the original line number from the compiled source list.
• domain: assigned by simple heuristic rules on the French side (presence of numerals, pronoun subjects, trailing ?, word count). These labels are approximate and should be treated as a rough filter, not ground truth.

No additional human annotation was performed for this initial release.

Who are the annotators?

Automatic only (heuristics). No human annotators.

Personal and Sensitive Information

None. The dataset contains only generic vocabulary and teaching phrases. One given name, "Numi" (a common Medumba name), appears in an example sentence about waking someone up; no real identifiable individual is referenced.

Bias, Risks, and Limitations

• Size. ~510 pairs is very small by MT-training standards. Treat this as an evaluation or fine-tuning seed, not a standalone training corpus.
• Domain skew. Heavy on everyday vocabulary, commands, body parts, and family terms; light on narrative, formal register, news, and technical content.
• Dialect variation unannotated. Medumba has regional variants (Bangangté, Bazou, Tonga, and more). The corpus likely leans Bangangté-general but does not mark locale.
• Tone fidelity. Tone marks are preserved as provided by the source material but have not been independently re-verified by a trained phonologist for every entry.
• Known typo. source_id=473 (the Medumba translation of "Il neige") contains cw:ǎd, almost certainly a typo for cwǎd given that cwɛ̌d is the form used elsewhere. Left as-is pending native-speaker review.
• No audio. Text only; tonal contrasts are underspecified by orthography alone.
• Ambiguous variants. Where one French source maps to several Medumba translations, all variants are marked as equally valid without register or pragmatic distinctions (e.g., short form vs. copula-full form).

Recommendations

• Do not use this corpus as the sole training source for a production system.
• Report results separately on test and document whether validation was used for hyperparameter tuning.
• For any downstream application, verify a sample of translations with a native Medumba speaker.
• When extending this corpus, preserve the source_id field and continue to split on unique French sources to avoid leakage.
• If releasing a derivative, keep the CC BY 4.0 attribution and acknowledge the Medumba-speaking community.

Citation [optional]

BibTeX:

@misc{french_medumba_parallel_2026,
  title        = {French--Medumba Parallel Corpus},
  author       = {Yannick KG},
  year         = {2026},
  howpublished = {Hugging Face Datasets},
  note         = {A seed parallel corpus for Medumba (byv), a Bamileke language of West Cameroon}
}

APA:

KG, Y. (2026). French–Medumba Parallel Corpus [Data set]. Hugging Face Datasets.

Glossary [optional]

• AGLC — Alphabet Général des Langues Camerounaises (General Alphabet of Cameroonian Languages). The standardized orthography used for Cameroonian national languages, including Medumba. Uses IPA-derived characters (ɛ, ɔ, ŋ, ʉ, ə) and explicit tone diacritics.
• Bamileke — A group of closely related Grassfields languages of western Cameroon; Medumba is one of them.
• Grassfields Bantu — The wider branch of the Niger-Congo family to which Medumba belongs.
• NFC — Normalization Form Canonical Composition, the Unicode normalization used throughout the dataset to ensure consistent byte-level representation of composed characters (e.g., tone-marked vowels).
• byv — ISO 639-3 code for Medumba.

More Information [optional]

If you extend this dataset, audit its translations with native speakers, or release a derivative, opening a pull request or discussion on the Hub repo is welcomed so that improvements can be shared with the community.

Dataset Card Authors [optional]

Yannick KG

Dataset Card Contact

Yannick KG — maseyan@yahoo.fr