Autonomous Hyperparameter Search Under a Fixed Five-Minute Budget on a Consumer GPU
We present a 35-run autonomous hyperparameter search campaign using the Windows/RTX fork of autoresearch on a single NVIDIA RTX 3080 (10 GB), trained on TinyStories under a fixed five-minute wall-clock budget per experiment, followed by a 17-run post-search confirmation sweep across 2–4 seeds for the most important configurations. Starting from an initial baseline of 1.027091 validation bits per byte (val_bpb), the search reached a best single-run result of 0.495600, a 51.7% improvement over the original baseline.
Key Result: 51.7% improvement over baseline (0.496 vs 1.027 val_bpb)