How men were left to starve in a South African gold mine. This image was created by Mohamed Hussein using the artificial intelligence (AI) tool Midjourney. Ayanda Ndabeni watched the faint glow from his headlamp fight the vast darkness 1,500 metres (4,920 feet) below ground. His miner's lamp had lasted for more than a week after he was lowered down into the shaft of the gold mine. But now the batteries were dying. He gently flipped the plastic switch of his lamp, turning it off, and the trapped men around him became shadows. In the stifling heat and humidity, their anxiety pressed in from all sides. Ayanda had descended into Shaft 10 of the Buffelsfontein mine in late September 2024, lowered by a team of nearly 20 men operating ropes and a pulley above ground. That day, he'd spotted police vehicles near the mine's entrance. The 36-year-old assumed it was just routine patrols around the mine system, which is 2km (1.2 miles) deep. But then the rope pulley, via which food, water, batteries and other items arrived, stopped moving. The shouting that usually indicated the rope operators were sending down a man or supplies also fell silent. When huge rocks came crashing down the shaft, they knew it was a warning. The men whispered of their growing fears that something was very wrong on the surface. Patrick Ntsokolo was also in Shaft 10. He was a few hundred metres higher up than Ayanda and had arrived in late July. Patrick was new to the mines. Tasked by the leaders of the artisanal miners with collecting the food, water and alcohol lowered down by the rope pulley, he hauled supplies along the slippery tunnels to small shops.

This paper describes the corrections made to the FLORES evaluation (dev and devtest) dataset for four African languages, namely Hausa, Northern Sotho (Sepedi), Xitsonga and isiZulu. The original dataset, though groundbreaking in its coverage of low-resource languages, exhibited various inconsistencies and inaccuracies in the reviewed languages that could potentially hinder the integrity of the evaluation of downstream tasks in natural language processing (NLP), especially machine translation. Through a meticulous review process by native speakers, several corrections were identified and implemented, improving the dataset's overall quality and reliability. For each language, we provide a concise summary of the errors encountered and corrected, and also present some statistical analysis that measure the difference between the existing and corrected datasets. We believe that our corrections enhance the linguistic accuracy and reliability of the data and, thereby, contributing to more effective evaluation of NLP tasks involving the four African languages.

When applied to processing long text, Large Language Models (LLMs) are limited by their context window. Existing efforts to address this limitation involve training specialized architectures, and cannot be easily applied to off-the-shelf LLMs. We present Parallel Context Windows (PCW), a method that alleviates the context window restriction for any off-the-shelf LLM without further training. The key to the approach is to carve a long context into chunks (``windows''), restrict the attention mechanism to apply only within each window, and re-use the positional embeddings across the windows. Our main results test the PCW approach on in-context learning with models that range in size between 750 million and 178 billion parameters, and show substantial improvements for tasks with diverse input and output spaces. We show additional benefits in other settings where long context windows may be beneficial: multi-hop questions and retrieval-augmented question answering with multiple retrieved documents. Our results highlight Parallel Context Windows as a promising method for applying off-the-shelf LLMs in a range of settings that require long text sequences. We make our code publicly available at https://github.com/ai21labs/parallel-context-windows.

In South Africa, it is a common practice for people to leave their vehicles beside the road when traveling long distances for a short comfort break. This practice might increase human encounters with wildlife, threatening their security and safety. Here we intend to create awareness about wildlife fencing, using drone technology and computer vision algorithms to recognize and detect wildlife fences and associated features. We collected data at Amakhala and Lalibela private game reserves in the Eastern Cape, South Africa. We used wildlife electric fence data containing single and double fences for the classification task. Additionally, we used aerial and still annotated images extracted from the drone and still cameras for the segmentation and detection tasks. The model training results from the drone camera outperformed those from the still camera. Generally, poor model performance is attributed to (1) over-decompression of images and (2) the ability of drone cameras to capture more details on images for the machine learning model to learn as compared to still cameras that capture only the front view of the wildlife fence. We argue that our model can be deployed on client-edge devices to inform people about the presence and significance of wildlife fencing, which minimizes human encounters with wildlife, thereby mitigating wildlife-vehicle collisions.