WeightsLab

Understanding AI models is not just about trust or robustness, it is about engineering speed and determinism. Small improvements in dataset and training quality can massively reduce iteration cycles; even a 1% data improvement translates into dramatically fewer experiments needed to reach target performance. WeightsLab helps ML teams inspect, debug, and optimize model training by exposing what is happening inside the training process. We currently focus on the data-facing side of perception systems: identifying problematic samples, dataset slices, annotation issues, redundancy, and training pathologies that aggregate metrics hide. Key capabilities include: Define custom training/debugging signals tailored to your task Query live training metrics and signals in natural language Create and manage dataset slices via manual tagging or natural-language bulk queries Filter, reweight, or discard data dynamically to reshape training distributions mid-experiment Our long-term vision is to become the operating system for understanding and optimizing the full AI training stack.