SalomonMetre13/nnb_fr_26k

Augmented Nande–French Parallel Corpus

This dataset is an augmented version of the Nande–French Parallel Corpus (nnb_fr_14k), containing 26,240 aligned Nande–French sentence pairs.

It was created to increase the size and linguistic variation of the original parallel corpus through French-side data augmentation using round-trip translation.

The resulting corpus is intended primarily for Nande–French and French–Nande machine translation research, particularly in low-resource language settings.

Dataset Summary

Property	Value
Languages	Nande (nnb), French (fr)
Language family	Bantu
Task	Machine Translation
Dataset type	Parallel Corpus
Number of sentence pairs	26,240
Train/Test split	90% / 10%
License	CC BY 4.0
Base dataset	SalomonMetre13/nnb_fr_14k
Augmentation	French-side round-trip translation
Augmentation models	OPUS-MT fr-en and en-fr

Dataset Structure

Each example consists of an aligned Nande–French sentence pair:

Field	Language	Description
nnb	Nande	Sentence in Nande
fr	French	Original or synthetically augmented French sentence

The corpus contains 26,240 aligned sentence pairs in total, divided into training and test sets using a 90:10 split.

Relationship to nnb_fr_14k

This dataset is directly derived from the SalomonMetre13/nnb_fr_14k dataset.

The nnb_fr_14k corpus contains 14,478 cleaned Nande–French sentence pairs derived from the original Nande–French corpus published through the Mozilla Data Collective.

For this release, the original corpus was further augmented by generating additional French variants for existing Nande–French sentence pairs.

Thus, this dataset should be viewed as an augmented extension of nnb_fr_14k, rather than as an independently collected parallel corpus.

Data Augmentation

The augmentation was performed on the French side of the original parallel corpus.

For each existing Nande–French pair, the original French sentence was processed using a French → English → French round-trip translation pipeline.

The following pretrained models were used:

• Helsinki-NLP/opus-mt-fr-en
• Helsinki-NLP/opus-mt-en-fr

The general augmentation process can be represented as:

Original Nande sentence
          │
          ▼
Original French translation
          │
          │ French → English
          ▼
      English text
          │
          │ English → French
          ▼
Synthetic French paraphrase

The resulting French sentences were paired with their corresponding Nande sentences and added to the original parallel corpus.

This procedure increases the number of available training examples while preserving the original Nande sentences.

Original vs. Synthetic Data

An important characteristic of this dataset is that it combines original parallel data with synthetically augmented parallel data.

The original French sentences originate from the base nnb_fr_14k dataset and ultimately from the original source corpus.

The additional French sentences were generated automatically using the Helsinki OPUS-MT French–English and English–French models.

Therefore, the augmented examples should not be considered independently human-translated sentence pairs.

The augmentation process is intended to introduce additional lexical and syntactic variation on the French side, but the generated sentences may not always be perfectly equivalent to their original counterparts.

Source and Provenance

The immediate source of this dataset is:

SalomonMetre13/nnb_fr_14k

The nnb_fr_14k dataset itself is a cleaned and prepared version of the Nande–French corpus originally published through the Mozilla Data Collective.

According to the original dataset description:

Translations were performed by professionals and volunteers of CLEAR Global's (formerly Translators without Borders) translator community.

The present dataset therefore has multiple stages of provenance:

Original Nande–French corpus │ ▼ Mozilla Data Collective │ ▼ Cleaned Nande–French corpus nnb_fr_14k (14,478 pairs) │ ▼ French-side augmentation via FR → EN → FR │ ▼ Augmented Nande–French corpus (26,240 pairs)

The augmentation process does not constitute additional human translation work.

Data Cleaning and Preparation

The starting point for this dataset is the cleaned nnb_fr_14k corpus.

The additional data was generated through the French-side round-trip translation process described above.

As a result, this release differs from both the original Mozilla Data Collective corpus and the nnb_fr_14k dataset.

Users interested in the original, non-augmented data should use the nnb_fr_14k dataset.

Intended Use

This dataset is intended for:

Nande–French machine translation French–Nande machine translation Fine-tuning neural machine translation models Data augmentation research Low-resource language NLP Multilingual and cross-lingual NLP Research on synthetic parallel data Development of language technologies for African languages Nande language processing

The dataset was used to fine-tune the pretrained T5-base model for the development of the following translation model:

SalomonMetre13/nnd_fr_mt_v3

Limitations

The synthetic portion of this dataset is generated through machine translation and should not be assumed to have the same quality or reliability as the original human-translated data.

The round-trip translation process can introduce:

Semantic drift: the generated sentence may differ slightly in meaning from the original. Information loss: details may be dropped during either translation step. Translation artifacts: generated French may contain unnatural or model-specific phrasing. Grammatical errors: the generated sentence may contain grammatical or stylistic issues. Repetition: the augmentation process may produce sentences that are structurally similar to existing examples. English-mediated effects: linguistic distinctions present in French may be altered as a result of passing through English.

These limitations should be considered when using the dataset for training or evaluation.

In particular, synthetically augmented examples should ideally not be used as the sole evaluation benchmark for machine translation systems.

Linguistic and Representation Considerations

Nande is a relatively low-resource language in natural language processing. The underlying corpus may not fully represent the diversity of Nande linguistic usage, including:

Regional and dialectal variation Different registers and writing styles Informal and conversational language Specialized domains The full range of Nande-speaking communities

The augmentation process operates on the French side only and does not introduce new independently translated Nande sentences.

Consequently, increasing the number of sentence pairs does not necessarily correspond to an equivalent increase in the linguistic coverage of Nande.

Recommended Use

The augmented examples are primarily intended to provide additional training data and target-side variation.

For model evaluation and benchmarking, it is recommended to use an evaluation set that is:

Independent of the augmentation process Human-translated or otherwise verified Not used during model training

When conducting experiments, researchers may also wish to distinguish between the original examples inherited from nnb_fr_14k and the synthetically generated examples.

Licensing

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

This dataset is derived from the nnb_fr_14k dataset, which itself is derived from the original Mozilla Data Collective resource.

Users should therefore review the licensing and attribution requirements of the underlying datasets.

The augmentation pipeline uses the following pretrained models:

Helsinki-NLP/opus-mt-fr-en Helsinki-NLP/opus-mt-en-fr

Users should consult the respective model cards and licenses for these models as well.

Acknowledgements

We acknowledge the contributors to the original Nande–French corpus, including the professional translators and volunteers of CLEAR Global's translator community (formerly Translators without Borders).

We also acknowledge the Mozilla Data Collective for hosting and distributing the original resource.

The present dataset builds directly upon the cleaned nnb_fr_14k dataset.

The synthetic augmentation was performed using the Helsinki-NLP OPUS-MT French–English and English–French translation models.

Related Resources

Base dataset: SalomonMetre13/nnb_fr_14k Translation model: SalomonMetre13/nnd_fr_mt_v3 French → English augmentation model: Helsinki-NLP/opus-mt-fr-en English → French augmentation model: Helsinki-NLP/opus-mt-en-fr Original source dataset: Mozilla Data Collective Citation

If you use this dataset in your research, please cite this dataset and acknowledge the underlying nnb_fr_14k corpus, the original source, and the contributors responsible for the human translations.

Base dataset: SalomonMetre13/nnb_fr_14k

Original resource: Mozilla Data Collective — Nande–French parallel corpus.

Translation contributors: CLEAR Global translator community (formerly Translators without Borders).

Augmentation models: Helsinki-NLP OPUS-MT fr-en and en-fr.