Latest

The integration of LLMs into the healthcare sector has become a natural response to the growing volume of medical information. Modern medicine generates a colossal amount of unstructured data, most of which is stored in text format. The ability of AI to effectively process, analyze, and systematize these arrays allows

Creating a universal autonomous driving system is impossible without considering the global context, as artificial intelligence algorithms directly depend on the diversity of input data. A model that demonstrates flawless performance on the perfectly marked highways of California can become dangerous in the snowy conditions of Scandinavia or the chaotic

Alignment of large language models (LLM) is performed through a complex RLHF pipeline. This approach is efficient, but is associated with high costs and technical difficulties. Direct Preference Optimization (DPO) offers a simple, transparent alternative for preference optimization and effective model alignment techniques. This approach preserves the advantages of RLHF,

Reinforcement Learning with Human Feedback (RLHF) is a key approach to aligning large language models (LLMs) with human values and expectations. Rather than relying solely on automated metrics, RLHF uses human ratings to train models to select relevant responses. This method enables LLMs to generate grammatically correct text and exhibit

Creating reliable autonomous driving systems requires colossal amounts of labeled data. However, collecting real-world information from roads has significant limitations. Real trips are expensive due to costs for fuel, equipment, and driver labor, and the annotation process takes months. Furthermore, it is nearly impossible to safely capture critical situations on

The stage of evaluating LLM responses is the most critical stage of development, as it is impossible to improve a product without an objective measurement of quality. However, with the evolution of technology, traditional verification methods have faced serious challenges that make them ineffective in today’s market conditions. Manual

Micromobility annotation is the labeling of cyclists, pedestrians, scooters, skateboards, bicycles, and other micro-mobility devices for training urban mobility AI in complex urban scenarios, including accurate e-scooter detection. The main goal is to transform complex street scenes into structured data through bounding boxes, keypoints, segmentation, and tracking, enabling reliable e-scooter