Diagnose a run · diagnose

Read the files of a training run and say whether it is healthy, and what to change if not:

adabmDCA diagnose model                                  # a finished or running run
adabmDCA diagnose model -s check/sampling.json           # with the check of the finished model
adabmDCA diagnose model -l run1 --json                   # a labelled run, full report as JSON

It reads history.csv, events.jsonl, training.json and the log of the run in the folder (with -l, the files prefixed by that label), and, with -s, the sampling.json written by adabmDCA sample --plot on the finished model. It prints a verdict and the findings behind it:

Verdict	Meaning
healthy	Nothing above an informational note
needs_attention	At least one warning, such as a run that hit its budget or a sampled Pearson well below the training one
failed	An error: the run stopped or stalled on trapped configurations
running	No summary yet and the files changed in the last 15 minutes
incomplete	No summary and no recent output: the run was cancelled or crashed

Each finding names the failure of the diagnostics guide it points to (F1–F8) and the change that guide recommends, as command-line options. The thresholds are the guide's heuristics, so the report also lists the indicators they were computed from.

From Python, adabmDCA.diagnose_run(run_dir, label=None, sampling_json=None) returns the same report as a dictionary.