README.md

---
language: nso
tags:
- xlm-roberta
- masked-language-model
- sepedi
- african-language
- sediba
- nlu
- nlp
license: mit
---

# sediba-XLM-R

A Sepedi masked-language model built on XLM-RoBERTa for Sepedi NLU research and token-level tasks.

## Model Details

### Model Description

- **Developed by:** Sediba AI NPC
- **Funded by:** Sediba AI NPC / TST (Pty) Ltd — Mankweng, Limpopo
- **Shared by:** Sediba AI NPC
- **Model type:** Masked language model (encoder-only Transformer)
- **Language(s) (NLP):** Sepedi (engpossible code-switched Sepedi/English contexts)
- **License:** MIT
- **Finetuned from model:** XLM-RoBERTa-base (Facebook/Google XLM-R)

### Model Sources

- **Repository:** huggingface.co/Sediba-AI/sediba-XLM-R
- **Demo:** `huggingface.co/spaces/Sediba-AI/leotsha` (7-mode CDI datafication game surface)
- **Paper:** [forthcoming]

## Uses

### Direct Use

- Masked-token prediction and probing for Sepedi
- Token-level NLU research (intent, NER probes) built on Sepedi representations
- Evaluation baseline for Sepedi-language NLP tasks

### Downstream Use

- Fine-tuning on Sepedi downstream tasks (classification, NER, etc.)
- Serving as the Sepedi representation anchor for the broader Sediba model family

### Out-of-Scope Use

- Generative/conversational Sepedi (this is an MLM, not a generative LM)
- High-stakes automated decisions about Sepedi speakers without human review

## Sepedi Families Context

This repo is the **Sediba Sepedi MLM anchor**. It is one model in a family with distinct roles:

|| Model | Role | Status | Repo / location |
||---|---|---|--|---|
|| **sediba-XLM-R** (this repo) | Masked-language Sepedi NLU anchor | **Shipped** | `Sediba-AI/sediba-XLM-R` on HF |
|| **Leotša la Sediba** | Conversational Sepedi (post-training / RLHF tier) | **Pending — publication holds until 1B+ base tier** | held in Sediba vaults; HF publication pending 1B+ base |
|| **SedibaLM V7** | QLoRA experiment: Qwen2.5-1.5B-Instruct + custom Sepedi vocab extension | **In training (Kaggle trail)** | Kaggle trail (`kaggle.com/sedibaai/sedibalm-v7-sepedi-qlora`) + local `v7_launch_bundle/` |

**Leotša la Sediba** and SedibaLM V7 are **not** published under this repo ID. They are tracked separately in Sediba's vaults and training pipeline, and differ from this model in base architecture, parameter tier, and task.

## Evaluation

- Self-language sanity probe via lglang on a held-out Sepedi sample
- Sepedi-width fertility check comparing model output diversity against a Sepedi baseline sample
- Exact results are tracked in the Sediba vaults (model evaluation log) and are available on request.

## Limitations

- Single-language Sepedi focus; performance on heavily code-switched or multi-register Sepedi is not fully characterised
- Masked-LM only — not a generative model; do not use for Sepedi text generation
- Training data scope and provenance are tracked in Sediba's data-governance stack (Esethu / NOODL / Itsebe) and the Sepedi corpus register

## Citation

If you use this model, cite Sediba AI NPC and link the repo:

```
@misc{sediba-xlmr,
  author = {Sediba AI NPC},
  title = {sediba-XLM-R: A Sepedi masked-language model built on XLM-RoBERTa},
  year = {2026},
  url = {https://huggingface.co/Sediba-AI/sediba-XLM-R}
}
```

Build by Sediba AI | Mankweng, Limpopo.