datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
Kuzi-Amharic-Uncensored-Datasetamharic-sentences-corpus
Amharic Sentences Corpus V1.0
Source: Telegram
This 1.6 million Amharic sentences corpus reflects current Amharic usage as of December 20, 2025, and is designed for anyone interested in:
Training Amharic-based LLMs
Fine-tuning NLP models
Building search, summarization, or generative systems in Amharic
The dataset is heavily cleaned and normalized, but like any serious LLM dataset, it still needs proper tokenization for pre-training.
I recommend using an… See the full description on the dataset page: https://huggingface.co/datasets/a3xrfgb/amharic-sentences-corpus.amharic-qa
AmQA: Amharic Question Answering Dataset
Amharic question and answer dataset in a prompt and completion format.
Dataset Details
In Amharic, interrogative sentences can be formulated using information-seeking pronouns like “ምን” (what), “መቼ” (when), “ማን” (who), “የት” (where), “የትኛው” (which), etc. and prepositional interrogative phrases like “ለምን” [ለ-ምን] (why), “በምን” [በ-ምን] (by what), etc. Besides, a verb phrase could be used to pose questions (Getahun 2013; Baye 2009). As… See the full description on the dataset page: https://huggingface.co/datasets/Henok/amharic-qa.adaption-amharic-text-corpus
This dataset is a remastered version of this dataset prepared using Adaption's Adaptive Data platform.
adaption-amharic_text_corpus
This dataset comprises over 700,000 Amharic text documents formatted as line-delimited JSON, covering diverse topics such as history, religion, politics, and product descriptions. Each entry contains a single string field with native Amharic content, including some samples with mixed languages or placeholder values. It is designed for text… See the full description on the dataset page: https://huggingface.co/datasets/Reubencf/adaption-amharic-text-corpus.amharic-alpacaMore details: https://arxiv.org/abs/2403.06354
amharic-dolly-15kAmharic version of the Dolly dataset (https://www.databricks.com/blog/2023/04/12/dolly-first-open-commercially-viable-instruction-tuned-llm)
Translated with this https://github.com/iocuydi/amharic-llama-llava/blob/main/data/prepare_amharic_data.py
More details: https://arxiv.org/abs/2403.06354
Amharic_corpus_split
Amharic Corpus — 4 x 5k Splits
A randomly shuffled subset of Reubencf/Amharic_corpus,
divided into four equal splits of 5,000 rows each (20,000 rows total).
Splits: split_1, split_2, split_3, split_4 (5,000 rows each)
Format: JSON Lines, one {"text": "..."} per line.
Sampling: random without replacement (seed 42); the four splits are mutually exclusive.
from datasets import load_dataset
ds = load_dataset("Reubencf/Amharic_corpus_split")
print(ds) # split_1..split_4, 5000 rows… See the full description on the dataset page: https://huggingface.co/datasets/Reubencf/Amharic_corpus_split.Amharic
Amharic Dataset (Cloned from saillab/taco-datasets)
This dataset contains the Amharic language part of the multilingual instruction tuning dataset, originally from the TACO dataset. The Amharic data is part of the multilingual-alpaca-52k-gpt-4 split, focusing on instruction-based dialogue data.
Dataset Summary
This dataset contains Amharic language instructions and dialogues. It is designed for fine-tuning large language models in multilingual settings. The dataset… See the full description on the dataset page: https://huggingface.co/datasets/CRLannister/Amharic.Amharic_corpus
Amharic Corpus (JSON)
Amharic text corpus, 707,649 rows, one document per row.
Format: JSON Lines (amh_data.jsonl) — one {"text": "..."} object per line.
Column: text (string)
Language: Amharic (am)
License: MIT
Originally distributed as a plain-text file; here it is provided as line-delimited JSON.
Usage
from datasets import load_dataset
ds = load_dataset("Reubencf/Amharic_corpus")
print(ds["train"][0]["text"])
amharic-blip-laionDataset used for pretraining clip alignment step of Amharic llava.
More details: https://arxiv.org/abs/2403.06354
amharic-property-mappingLugha
Lugha – Hausa Speech Dataset
Crowd-sourced Hausa voice recordings collected via the Lugha mobile app.
Structure
audio/sample_XXXX.m4a – raw audio (m4a)
metadata.jsonl – one JSON object per recording
Fields
field
description
audio
relative path to the audio file
text
prompt sentence read by the speaker
language
spoken language
state
Nigerian state of the speaker
lga
Local Government Area
accent
self-reported accent
age_range… See the full description on the dataset page: https://huggingface.co/datasets/Amhaztech/Lugha.new-amharic-passage-retrieval-dataset
New Amharic Passage Retrieval Dataset (AMQA)
This dataset is generated from AMQA-style question–context pairs, converted to match
the schema of Desalegnn/amharic-passage-retrieval-dataset.
Columns
query_id, passage_id, query, passage, category, link
Load Example
from datasets import load_dataset
url = 'https://huggingface.co/datasets/Desalegnn/new-amharic-passage-retrieval-dataset/resolve/main/train.jsonl'
ds = load_dataset('json', data_files={'train': url})… See the full description on the dataset page: https://huggingface.co/datasets/Desalegnn/new-amharic-passage-retrieval-dataset.gamma-g1-327-vast-g1-326-amharic-safety-gate-20260624eng-to-amh-translationmtruthfulqa-amh_Ethigamma-g1-325-amharic-safety-closure-repair-data-20260623gamma-g1-328-vast-g1-326-amharic-decode-boundary-gate-20260624amh_Ethigetedil-amharic-datasetAmharic
Amharic Dataset (Cloned from saillab/taco-datasets)
This dataset contains the Amharic language part of the multilingual instruction tuning dataset, originally from the TACO dataset. The Amharic data is part of the multilingual-alpaca-52k-gpt-4 split, focusing on instruction-based dialogue data.
Dataset Summary
This dataset contains Amharic language instructions and dialogues. It is designed for fine-tuning large language models in multilingual settings. The… See the full description on the dataset page: https://huggingface.co/datasets/yohannabelay/Amharic.amharic-wikipediaamharic_primary_school_examsspp_v1SPP v1, mesh and segmentations.
amharic-geez-numerical-blindspot
Blind Spots of Frontier Models: Tokenizer-Induced Numerical & Temporal Collapse in Ge'ez and Amharic
Author: Teshome Birhanu Cheru
Affiliation: Addis Ababa University, Electrical and Computer Engineering
Target Fellowship: Fatima Fellowship 2026 Technical Challenge
Evaluated Model: Qwen/Qwen2.5-3B-Instruct (3B Parameters)
Artifacts Repository: Evaluation Benchmark (benchmark_ethiopian_reasoning.json), Empirical Outputs (eval_results.json), Colab Notebook… See the full description on the dataset page: https://huggingface.co/datasets/abute-21/amharic-geez-numerical-blindspot.