Datasets:
English-Swahili Agricultural Parallel Corpus (MT-Eng-Swa)
Dataset Summary
MT-Eng-Swa is a parallel corpus of 52,170 English-Swahili sentence pairs in the agricultural domain. It is published by AI4Agric to support machine translation and agricultural advisory services for Swahili-speaking farmers in Kenya and the wider East African region. The English side consists of farmer questions and extension content covering crops, livestock, soil fertility, pest management, post-harvest handling and agribusiness. Each English sentence has a human translation into Swahili that has been reviewed and approved.
The corpus is one of three parallel datasets released together. The other two are MT-Eng-Kik (Kikuyu) and MT-Eng-Kln (Kalenjin). Around 46,500 serial numbers are shared across all three datasets. This allows multi-way alignment through English, as well as Swahili-Kikuyu and Swahili-Kalenjin pairs to be built through the shared identifiers.
Languages
| Role | Language | ISO 639 code |
|---|---|---|
| Source | English | en |
| Target | Swahili (Kiswahili) | sw / swa |
The Swahili text follows the variety used in Kenya. Swahili is a Bantu language and a national and official language of Kenya.
Dataset Structure
Data Fields
| Field | Type | Description |
|---|---|---|
source_text |
string | English sentence |
target_text |
string | Swahili translation |
language |
string | Target language (always swahili) |
serial_number |
string | Unique identifier for the sentence pair. Prefixes are Hq or KARLO |
import_source |
string | Origin of the English sentence (Huggingface, KARLO or Farmerchat) |
Data Instance
{
"source_text": "Advantage of feeding packchong my cow",
"target_text": "Faida ya kulisha ng'ombe wangu nyasi aina ya Packchong",
"language": "swahili",
"serial_number": "Hq007",
"import_source": "Huggingface"
}
Data Splits
The dataset is released as a single train split of 52,170 pairs. No official validation or test split is defined. Users who report results should state how they split the data and should fix a random seed.
Dataset Statistics
Source Composition
import_source |
Pairs | Share |
|---|---|---|
| Huggingface | 46,285 | 88.7% |
| KARLO | 3,555 | 6.8% |
| Farmerchat | 2,330 | 4.5% |
| Total | 52,170 | 100% |
Sentence Length (words)
| Mean | Median | Min | Max | Total tokens | Vocabulary (lower-cased) | |
|---|---|---|---|---|---|---|
| English | 11.0 | 11 | 4 | 108 | 571,431 | 17,450 |
| Swahili | 12.5 | 12 | 1 | 158 | 654,100 | 22,709 |
Swahili translations are on average slightly longer than their English sources. The Swahili vocabulary is larger than the English one, which reflects the noun class prefixes and verbal affixes of Swahili morphology.
Usage
from datasets import load_dataset
ds = load_dataset("AI4Agric/MT-Eng-Swa", split="train")
print(ds[0])
# Create a reproducible 80/10/10 split
tmp = ds.train_test_split(test_size=0.2, seed=42)
vt = tmp["test"].train_test_split(test_size=0.5, seed=42)
train, validation, test = tmp["train"], vt["train"], vt["test"]
Dataset Creation
Source Data
The English sentences come from three sources:
- Huggingface (
Hqserial numbers): farmer questions from an existing agricultural question dataset on Hugging Face. TODO: name and link the dataset. - KARLO (
KARLOserial numbers): sentences from agricultural extension and research material. TODO: give the full name of the source organisation. - Farmerchat (
Hqserial numbers): farmer queries from the FarmerChat advisory service. TODO: confirm and link.
Many of the Huggingface and Farmerchat sentences are real farmer questions. They keep informal spelling and phrasing (for example, "Advantage of feeding packchong my cow"). These have been left as they are because they reflect how farmers actually write. The Swahili translations are written in standard Swahili. Product and variety names such as Packchong and T5Z are kept in their original form.
Translation and Review
Native Swahili speakers translated the English sentences. Each translation was then reviewed and approved before it was included. TODO: describe the number of translators, how they were recruited, the review protocol and the platform used.
Cleaning Applied Before Release
Leading, trailing and repeated whitespace was removed. Pairs with an empty source or target were dropped (one pair). Exact duplicate pairs were removed, which left 52,170 pairs out of 52,171.
Considerations and Limitations
- Domain. The corpus is specific to agriculture. Models trained only on it will not generalise well to other domains.
- Repeated source sentences. Three English sentences occur more than once with different Swahili translations. These should be kept in the same split to avoid leakage.
- Very short targets. Some Swahili translations are a single word. Users may want to filter pairs with extreme length ratios.
- Noisy source text. Farmer questions contain spelling and grammar errors on the English side.
- Mixed apostrophes. Words such as ng'ombe appear with both straight (') and curly (’) apostrophes. Normalising them before training is recommended.
Related Datasets
- AI4Agric/MT-Eng-Kik: English-Kikuyu agricultural parallel corpus
- AI4Agric/MT-Eng-Kln: English-Kalenjin agricultural parallel corpus
Citation
@misc{ai4agric_mt_eng_swa_2026,
author = {AI4Agric},
title = {{MT-Eng-Swa}: English-Swahili Agricultural Parallel Corpus},
year = {2026},
publisher = {Hugging Face},
howpublished = {\url{https://huggingface.co/datasets/AI4Agric/MT-Eng-Swa}}
}
- Downloads last month
- 36