Gap report
What Ethiopian AI is still missing
Every claim on this page is derived from the index — the same daily scrape that powers the rest of goha. Numbers are counts of what is publicly published, not estimates. If something is missing here, it is missing on Hugging Face, GitHub and arXiv too.
7 languages · 6 tasks · 42 cells analysed
16
empty cells
of 42 language × task combinations have no model, dataset or paper.
172
unevaluated models
of 189 models publish no evaluation result.
312
unlicensed assets
models and datasets that declare no license — unclear for reuse.
1
research-only cells
tasks with published papers but no model or dataset to use.
Coverage by language
Models + datasets per task. A dash means nothing is indexed; a number in parentheses is published papers.
| Language | ASR | TTS | Translation | LLM | NER | Embeddings |
|---|---|---|---|---|---|---|
| Amharic | 20(5) | 15 | 5(24) | 19(15) | 3 | 9 |
| Tigrinya | 16(2) | 3(1) | 2(7) | 13(6) | 2 | — |
| Afaan Oromo | 10(2) | 4 | 1(2) | 3(3) | 3 | — |
| English | 2(1) | — | 7(19) | 11(5) | 2 | 2 |
| Somali | — | — | —(1) | 1(2) | — | — |
| Wolayttattoo | 3(1) | — | — | — | — | — |
| Sidaamu Afoo | 3 | 2 | — | — | — | — |
Nothing exists at all
Language × task combinations with no model, dataset or paper — the blank spots on the map.
- TigrinyaEmbeddings
- Afaan OromoEmbeddings
- EnglishTTS
- SomaliASRTTSNEREmbeddings
- WolayttattooTTSTranslationLLMNEREmbeddings
- Sidaamu AfooTranslationLLMNEREmbeddings
Research leads, assets lag
Tasks with published papers but no model or dataset to actually use — knowledge exists, tooling does not.
- Somali · Translationcoverage →
Popular but unmeasured
The most-downloaded models that publish no evaluation result. Usage without evidence is the ecosystem's quietest risk.
- wav2vec2-large-xlsr-53-amharicagkphysics73.9k/mo
- Llama-3.2-1B-Amharic-Instructrasyosef1.2k/mo
- Ethio-ASR-amharicbadrex994/mo
- roberta-base-amharicrasyosef532/mo
- amharic-captions-modelkal11529/mo
- Llama-3.2-1B-Amharicrasyosef485/mo
- omnivoice-amharicLab-et465/mo
- bert-base-multilingual-cased-finetuned-amharicDavlan429/mo
Unclear licensing
312 assets declare no license. For companies building on public models, that is a legal grey zone — reuse is technically possible, commercially risky.
Methodology
Languages and tasks are the ones the index tracks (ASR, TTS, translation, LLM, NER, embeddings). A cell counts a model or dataset when the asset declares the language and task in its metadata. Papers are matched by their derived language/task tags. "No evaluation" means the model card publishes no metrics table. "No license" means the card declares none. Everything is recomputed on every daily scrape — numbers here move as the ecosystem moves.
Related: language coverage · published benchmarks · ask about a specific gap
