autoresearch

Automating hyperparameter tuning for single-GPU ML experiments requires running a fixed 5-minute training budget on `train.py` and evaluating validation bpb to determine if changes should be kept. This approach uses branch-based iteration to advance improvements and discard runs that do not enhance the target metric.

Branch-based iteration can manage ML architecture search by creating dedicated experiment branches tagged by date, running training, and resetting or discarding branches when results fail to improve validation bpb. This mechanism ensures only beneficial architecture modifications are retained in the workflow.

Optimizing val_bpb during short ML research cycles is best achieved through autonomous metric-first decision making that logs results strictly to `results.tsv`. The system evaluates ground-truth metrics after a deterministic wall-clock budget and keeps only changes that yield the lowest validation bpb.

A fixed evaluation harness is required to run autonomous ML research experiments because the system minimizes validation bpb against a consistent baseline. It enforces strict edit constraints on in-scope files and logs peak memory fields alongside the target metric for accurate comparisons.

A 5-minute training budget limits ML experimentation to short, single-GPU research cycles where rapid feedback is prioritized over long-running convergence. It applies strictly to scenarios where only training code is editable and requires strict edit constraints on in-scope files to prevent uncontrolled modifications.