Sediba-AI/sepedi-training-v1

Sediba Sepedi Training Corpora v1

Sepedi (Sesotho sa Leboa) training data assembled by Sediba AI, Mankweng, Limpopo, South Africa.

Sepedi is spoken by roughly 4.7 million people and is severely under-represented in language models. This release is one of the larger openly-listed Sepedi training collections.

Contents

From MANIFEST.json (v1.0, 2026-06-05) — 884,987 records, ~289 MB:

File	Records	Purpose
mt560_clean.jsonl	397,861	Translation pairs
lm_train.jsonl	376,803	Language-model pre-training text
instruction_all.jsonl	41,687	Instruction-following (superset)
alpaca_instruction_pairs.jsonl	41,601	Alpaca-style instruction pairs
lm_eval.jsonl	19,832	LM evaluation split
sepedi_ner.jsonl	7,117	Named-entity recognition
instruction_pairs.jsonl	86	Hand-written instruction seeds
leotsha_labels.json	—	Sentiment labels (Leotša platform)

Note instruction_all.jsonl (41,687) largely overlaps alpaca_instruction_pairs.jsonl (41,601) — it is a superset, not an additional 41k records. Deduplicate before combining.

Usage

from datasets import load_dataset

# language-model pre-training text
ds = load_dataset("Sediba-AI/sepedi-training-v1", data_files="lm_train.jsonl")

# instruction tuning
ds = load_dataset("Sediba-AI/sepedi-training-v1", data_files="instruction_all.jsonl")

Models trained on this data

• Sedibaai/SedibaLM — generative Sepedi model (Qwen2.5-1.5B + QLoRA), validated against its base
• Sediba-AI/sediba-XLM-R — Sepedi masked-LM
• Sediba-AI/xlmr-sepedi — Sepedi masked-LM

⚠️ Limitations — read before using

• No formal data-licence audit has been completed. Component corpora were assembled from multiple sources; per-source licence provenance is not fully documented in this release. If you intend commercial use, audit provenance yourself first. This is stated plainly because an undocumented provenance claim is worse than an acknowledged gap.
• mt560_clean.jsonl derives from the MT560 multilingual translation collection, which carries its own upstream terms — check them independently.
• Quality is not uniform. The collection mixes curated, scraped, and synthetic/augmented material. There is no per-record quality score.
• Dialect distribution undocumented. Sepedi varies substantially across the five Sediba dialect zones; this release carries no zone stratification.
• Not deduplicated across files (see the instruction-file overlap above).
• Some records may be English or code-switched, particularly in translation pairs.
• No personal-information audit. Do not assume the text is free of identifiers.

Licence

CC BY-NC-SA 4.0 — attribution, non-commercial, share-alike.

This is the interim licence pending legal review of NOODL (Northern-Sotho Open Data Licence), per the Sediba data-governance framework. Attribution must name the contributing communities, not only "Sediba AI".

Commercial use requires a separate agreement. Component corpora may carry their own upstream terms that this licence cannot override — see Limitations.

Note that models trained on this data are released under more permissive terms (SedibaLM is CC BY-SA 4.0): open weights, restricted corpus.

Provenance and consent

Community-contributed material was gathered under FPIC (Free, Prior and Informed Consent) principles, with consent recorded per contributor and tagged by dialect zone. The Kutullo Exchange royalty mechanism is designed so contributing communities share in downstream value.

Do not use this data to build systems that mislead or harm Sepedi-speaking communities, or for surveillance, social control, or voter manipulation.

Citation

@misc{sediba-sepedi-corpora-v1,
  title  = {Sediba Sepedi Training Corpora v1},
  author = {Sediba AI},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Sediba-AI/sepedi-training-v1}
}

About

Sediba AI — sovereign AI for South African languages. Sepedi first, then the remaining official languages of South Africa, then the wider Global South.