Lacuna Fund - Open Data Sets & Use Cases

This speech dataset for the Ghanian languages Akan (Akuapem Twi, Asante Twi, Fante) and Ga includes 104,000 utterances (speech) across the four dialects/languages with approximately 200 speakers per dialect/language. This amounts to about 148 hours of speech in total. The dataset was developed to support the development of financial applications in native Ghanaian languages to allow illiterate and semi-literate people to fully benefit from digital financial services. Secondly, it aims to answer research questions related to domain-specific vs. general-purpose dataset development, dialects, as well as NLP system development in low resource settings. Overall, a total of 83,829 audios were recorded from which the datasets were published and made publicly accessible.

The dataset was created to enable translation from Kiswahili, which is the national language in Kenya, into three indigenous languages, namely, Kidaw'ida, Kalenjin, and Dholuo. All three languages are low resource, especially Kidaw'ida, which has only around 400,000 speakers and is at immediate risk of loss. By collecting the text and speech data, the project has taken a step towards preservation of these languages . The dataset consists of three parallel corpora: Kidaw'ida-Kiswahili; Kalenjin-Kiswahili; Dholuo-Kiswahili. On average, each corpus has thirty thousand sentence pairs. In addition, the dataset has helped to preserve, revitalise and elevate the languages Kidaw'ida, Kalenjin, and Dholuo by making them work on Mozillas Common Voice Platform. The text and speech datasets are thereby supporting crowd-sourced voice recognition, promoting linguistic diversity and empoweromg local communities by enabling Natural Language Processing applications tailored to their needs. Namely, the resulting voice recognition datasets can be used to build AI voice applications such as advisory systems in local languages for farmers, citizens etc that "understand" speech input (instead of written input). As of August 2025, 120 hours of speech data have been recorded on Common Voice in Dholuo (with 14692 sentences available), 92 hours have been recorded for Kalendjin (With 29900 sentences available) and 56 hours have been recorded for Kidaw'ida (with 11773 sentences available). To download the crowd-sourced speech datasets in Kidaw'ida, Kalenjin, and Dholuo, visit https://commonvoice.mozilla.org/en/datasets

This dataset will help data scientists, government and users to measure solar energy adoption across Madagascar. It laid the groundwork needed to develop a solar panel detection algorithm working in Madagaskar. Notably, this project represented all regions of the country; instead of focusing only on big cities, it also covered average and small villages as well as coasts and mountains. The team annotated 2,125 Google Earth satellite images and 9,202 drone images, forming a combination of low and high-definition solar panel views in Madagascar. The Madagascar Initiatives for Digital Innovation (MAIDI) team performed field checks for up to 25% of satellite images and, in total, annotated 22,488 polygons.

This dataset supports yield estimation, crop type detection and classification, fruit detection and counting, and fruit maturity stage detection (unripe, ripe, and spoiled) for three products that are important sources of livelihood for millions of households in Sub-Saharan Africa. It contains 14,870 drone images with bounding box annotations of cashew, cocoa, and coffee trees collected across multiple farms in Ghana and Uganda. Conventional methods of yield estimation are expensive, require a lot of labor and time, and are prone to error due to incomplete ground observations. This results in poor crop yield estimations and hinders farmers’ ability to appropriately plan and manage their fields and production pipelines. This dataset will help transform African agriculture into agribusiness by allowing for the development of yield estimation solutions that enable farmers to make good business decisions. Having key details about agricultural production readily accessible enables a timely harvest, helping farmers ensure healthy, fresh produce and, in addition, better sales.

The Lacuna PII Multilingual Text Dataset contains annotated sentences with personally identifiable information (PII) in Luganda, Lumasaba, Hausa, and Kanuri. These four languages span Central and Eastern Uganda, Nigeria, Ghana, and Northern Cameroon. The dataset can help to anonymize and pseudonymize personally identifiable information in AI tasks. This can help organizations comply with legal requirements while still being able to analyze and use and share their data effectively. It can also be used to improve machine translation systems for low-resource languages, improve the performance of NLP applications in these languages, and support the extraction of specific information from text, such as automated form filling and information retrieval systems. It comprises 4000 Luganda sentences, 5000 Lumasaba sentences, 3000 Kanuri sentences, and 3000 Hausa sentences. The team aimed to curate a dataset that is gender inclusive. It was created by Makerere Artificial Intelligence Lab in collaboration with Marconi Lab and Clear Global.

This dataset helps to understand energy consumption patterns in relation to weather conditions in Pakistan. This can guide policymaking on energy and energy conservation, sustainabe energy initiatives and private sector use (market assessment and strategic planning as new entrants in the restructured energy market). The project produced an energy dataset of residential consumption data from buildings across six climatic zones. The energy dataset is recorded on a 1-minute granularity for entire household usage as well as for individual appliances. With over a year of detailed energy consumption data and 18 months of weather data collected from a diverse array of households across six distinct climatic zones, it is one of the most comprehensive datasets of its kind. This meteorological dataset has been accumulated over a period of 18 months for the six urban centers. This weather data comprises of up to 10 essential variables such as temperature, atmospheric pressure, humidity and Precipitation, thus providing an all-around perspective of the environmental elements impacting energy consumption.

The team developed a labeled training dataset, derived from 50cm or better satellite imagery, based on a novel, pre-defined road space classification taxonomy appropriate for training and deployment of large-scale deep-learning models

The dataset can help to build systems, that can monitor the impact of palm oil monoculture, shrimp aquaculture, mining and other land transformations in continental Ecuador and Galapagos. The project created a 20.000 points land use/cover classification training dataset from existing data, with labels that can be used to train multi-spectral Earth observation (EO) data machine learning (ML) models covering continental Ecuador and the Galapagos islands.

This dataset was produced by the Centre for Remote Sensing and Geographic Information Services (CERSGIS) as part of the project Reference Data Collection for Improving Land Use Change Mapping in Ghana. The primary objective was to develop high-quality reference data to enhance the accuracy of remote sensing-based land use and land cover (LULC) change mapping using machine learning methods in Ghana’s cocoa production landscapes.

This project developed a efficient AI-powered method and dataset for monitoring of trees in forests in Kenya. Accurate tree crown detection and other parameters will support reforestation efforts in several ways: The datasets enable the development of tools that can rapidly assess the success of reforestation efforts by enabling counting of trees, estimation of trees’ biophysical parameters from drone images, and prediction of tree growth in relation to tree species and weather variables. The data consists of a high-resolution dataset of drone imagery, ground truth tree parameter measurements, and aggregate weather data from a large, reforested area in Kenya.

Deforestation is a significant contributor to changes in land use/cover in Uganda. This challenge is intensified by agriculture’s extreme vulnerability to climate change. This project has generated datasets that can be utilized to predict land cover types or changes for informed decision-making in natural resource management. The data is accessible, quality satellite imagery datasets and National Forestry Inventory datasets. Specifically, Sentinel 2 and Hansen tree cover image datasets have been augmented with the inclusion of inventory datasets. A biomass mean stock per hectare dataset for each vegetation type has also been acquired. Annotation procedures of the different datasets allow for machine learning tasks.

Extensive bioinformatics resource that leverages tree species’ medicinal, food provision, and other trait data, together with southern African trees’ climate relationships and growth characteristics for climate adaptation and mitigation planning. The data can serve to promote the use of indigenous trees for reforestation, regenerative agriculture, ecological restoration, human health and livelihood support, and urban afforestation programs to adapt to and mitigate the impacts of climate change.

Mangroves are valuable coastal ecosystems, providing diverse services to humans and playing a critical role in the global strategy to mitigate impacts of climate change. Regular monitoring of their conditions is essential to provide evidence that supports timely decisions for their management. The “Benin Mangroves Open Data (BeMOD)” project generated annotated images, observations on water quality, observations on soil characteristics, climate observations, forest inventories data (tree diversity, canopy height and cover, tree diameter, total height, biomass, etc.) from plots, and data on socio-cultural contexts (e.g., uses, activities) of mangroves in multiple sites in Benin.

This dataset contributes to improved understanding and mitigation of the impacts of oil palm cultivation on forests and climate change. It can also serve to support sustainable forest management practices and empower local communities in forest resource management. The project created open and accessible training and evaluation datasets for machine learning applications focused on forest cover and change in Indonesia. It contains collections of polygons of uniform land cover, with labels indicating land cover type and, possibly, information on land cover change over time. To label land cover patches, the team has reviewed high-resolution satellite imagery facilitated by Collect Earth Online (CEO) and augmented it with field visits to difficult-to-classify areas. It has prioritized areas with significant landcover diversity with a preference for areas with significant smallholder production of oil palm, coconut, and other commodities - given the well-known difficulties of distinguishing oil palm from coconut palm with lower resolution imagery. Possible users of these outputs are local communities engaged in community forestry in Indonesia, as well as researchers, policymakers, and civil society organizations. The dataset also contributes to social forestry, which recognizes that people who depend on local forests are best placed to look after them, and that allowing communities to manage and use forest resources can have positive social, environmental and economic impacts. Social forestry is a broad term and has different names in different places. Some examples are community forestry, village forestry, participatory forestry, community-based forest management and people-centred forestry.