Open models and datasets
for African languages.
The same research that powers our API is open. We release Amharic and Afan Oromo models and datasets on Hugging Face so the community can build on them, audit them, and push the languages forward with us.
Fine-tuned for the languages.
Continued pre-training, instruction tuning, and speech recognition. The open checkpoints behind our text-to-speech, speech-to-text, and language model APIs.
Language models
Amharic large language models, from continued pre-training through instruction tuning.
Gemma 2 27B 路 Amharic CPT
b1n1yam/gemma-2-27b-amharic-cptGemma 2 27B continued pre-trained on a large Amharic corpus. The base for our Amharic reasoning stack.
Gemma 2 27B 路 Amharic SFT
b1n1yam/gemma-2-27b-amharic-alpaca-sftThe CPT base supervised fine-tuned on Amharic Alpaca instruction data for chat and task following.
Qwen 路 Amharic 2B
b1n1yam/qwen-amh-2bA compact 2B Amharic text-generation model for lighter, lower-cost workloads.
Speech recognition
The Shook family of Amharic speech-to-text models, from on-device tiny to a 0.8B flagship.
Shook Medium 路 Amharic 2k
b1n1yam/shook-medium-amharic-2kOur flagship open Amharic speech-to-text model, trained on roughly 2,000 hours of audio.
Shook Tiny 路 Amharic 600hr
b1n1yam/shook-tiny-amharic-600hrA tiny Amharic ASR model trained on 600 hours, built for on-device and low-latency use.
Shook Tiny 路 Stage 2 Polish
b1n1yam/shook-tiny-amharic-stage2-polishA polished tiny Amharic ASR checkpoint with a second-stage refinement pass.
The data the languages were missing.
Text corpora, transcribed speech, and instruction sets for Amharic and Afan Oromo, including some of the largest open datasets for these languages anywhere.
Text corpora
Large pre-training corpora for Amharic, including one of the largest open Amharic text datasets on Hugging Face.
Amharic Combined Corpus
b1n1yam/amharic-combined-corpusOur largest open Amharic text corpus, combining web, encyclopedic, and curated sources for pre-training.
Amharic Corpus TG
b1n1yam/amharic_courpus_tg_1A large Amharic text corpus assembled for language-model pre-training.
Wikipedia Amharic
addisai/wikipedia-amharicA cleaned Amharic Wikipedia dump, ready for pre-training and retrieval.
Speech and audio
Transcribed speech for Amharic and Afan Oromo, powering both ASR and TTS research.
Waxal Afan Oromo TTS
b1n1yam/waxal-orm-tts-mergedA merged Afan Oromo text-to-speech dataset for voice synthesis research.
Sagalee
b1n1yam/SagaleeAn Afan Oromo speech corpus for training and evaluating speech recognition.
Afaan Oromo TTS
b1n1yam/afaan-oromo-ttsAn Afan Oromo text-to-speech dataset for voice synthesis.
WXL Amharic
b1n1yam/wxl_amhAn Amharic speech dataset for ASR and TTS work.
Instruction tuning
Amharic instruction datasets we released to make fine-tuning open models easier.
Alpaca Amharic
addisai/FreedomIntelligence-alpaca-amharicAn Amharic translation of the Alpaca instruction set for supervised fine-tuning.
FineTome Amharic
addisai/FineTome-single-turn-dedup-amharicA deduplicated single-turn Amharic instruction dataset for instruction tuning.
Build on the open stack.
Browse every model and dataset on Hugging Face, or start with the production API.