Dataset Viewer

The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.

Couldn't infer the same data file format for all splits. Got {NamedSplit('train'): (None, {}), NamedSplit('test'): ('json', {})}

Error code:   FileFormatMismatchBetweenSplitsError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Munyarwanda AI — Data Pipeline

Production data acquisition, cleaning, validation, deduplication, mixing and training pipeline for Munyarwanda AI, a Kinyarwanda-first LLM.

Dataset: https://huggingface.co/datasets/arcange9/Munyarwanda-AI-Dataset (creates config v0.3)
Model target: https://huggingface.co/arcange9/Munyarwanda-AI-v0.3
Colab notebook: notebooks/Munyarwanda-AI-v0.3-Full-Data-Pipeline.ipynb

What's here

data/
  resources_manifest.json   43 verified resources + licenses (research-backed)
  raw_manifests/           download manifests (raw data never committed)
  processed/               cleaned → dedup → final datasets (runtime output)
  reports/                 discovery / validation / language / dedup / mixture / bench reports
configs/
  data_mix_v0.3.yaml       mixture targets & ratios (rationale documented inline)
  training_v0.3.yaml       explicit training hyperparameters (QLoRA on Qwen3-0.6B)
scripts/
  discover_resources.py    re-verify manifest against live APIs
  download_datasets.py     fetch/stream approved datasets
  validate_datasets.py     schema + license gate
  common.py                Kinyarwanda language detector + quality scoring
  clean_text.py            normalization, language ID, provenance
  language_filter.py       language distribution audit
  deduplicate.py           exact hash + MinHash LSH (cross-dataset)
  build_pretraining_corpus.py / build_instruction_data.py /
  build_translation_data.py / build_eval_data.py / build_dataset_mix.py
notebooks/
  Munyarwanda-AI-v0.3-Full-Data-Pipeline.ipynb   29-step end-to-end run
docs/
  KINYARWANDA_DATASET_CATALOG.md   full 43-resource catalog w/ licenses & tiers
  DATA_LICENSES.md                  license policy + attribution block
  DATA_PIPELINE.md                 architecture walkthrough
  DATASET_REPORT.md                report template (auto-filled per run)
  MUNYARWANDA_BENCH.md             evaluation suite documentation

Quickstart (free Google Colab T4)

Open notebooks/Munyarwanda-AI-v0.3-Full-Data-Pipeline.ipynb in Colab.
Add a Colab secret named HF_TOKEN (Hugging Face token).
Accept the gated-dataset conditions once for mbazaNLP/kinyarwanda_monolingual_v01.1.
Set MODE = "SMALL" (test) or "FULL", then Run all.

The pipeline adapts to the detected GPU (VRAM, BF16/FP16) automatically.

Principles

Quality × diversity × correctness × license-safety × provenance × dedup over raw size (a smaller clean corpus beats a huge corrupted one).
No raw data in git — manifests and downloaders only.
Every trained example carries provenance (source, license, category, score).
Evaluation data is never trained on.
Nothing is used because it is merely downloadable; licenses are verified.

Research summary (2026-09-04)

43 resources verified via live Hugging Face + GitHub API queries: 26 accepted for training (12 SAFE_FOR_TRAINING, 14 TRAINING_WITH_ATTRIBUTION), 9 UNCLEAR (train but never redistribute), 5 RESEARCH_ONLY, 3 DO_NOT_USE (empty repos / personal chat PII). Full details: docs/KINYARWANDA_DATASET_CATALOG.md.

Downloads last month: 56

Total file size:

198 kB