Bogdan Gaza explains DatologyAI’s approach to synthetic training dataSynthetic data is information generated to represent examples or situations rather than collected directly from the events it describes., using selected source documents as seeds for rephrasing and restructuring. He presents the company’s reported comparisons as its own results.
He describes moving research and production workloads toward shared Kubernetes-based orchestration, then addresses bottlenecks in object-storageObject storage keeps data as independently addressed objects with payloads and metadata in a flat, scalable namespace. metadata and recovery from GPU failures. Batching metadata requests, partition sizing and checkpointsCheckpointing saves a defined execution state so a computation or workflow can resume or recover from that point. helped reduce wasted time in the reported workflow.
Gaza discusses coordinating CPU and GPU resourcesResource utilization measures how much of an available computing resource is in use during a defined period or workload. across clusters and tuning inferenceInference is the use of a trained model to produce an output from supplied input. parameters through experiments. He reports that the system scaled from smaller runs to a 12-trillion-token generation effort across web, math and code.
Watch on YouTube
