Dora manages ML experiment configurations via content-addressed signatures rather than manual naming

Demucs training uses Dora (a Meta experiment-management library) to track hyperparameter configurations. Each unique set of hyperparameters produces a deterministic hash (signature), e.g. 9357e12e. The hash is computed as a delta from a base config — this means the base conf/*.yaml files must never be directly edited for hyperparameters; only overrides passed on the command line produce correct signatures. If you edit the base config, stored checkpoints can no longer be correctly matched to their configurations. This is a form of configuration-as-content-address: the signature uniquely identifies the exact training run, enabling reproducibility without manual versioning.

Examples

dora info -f 9357e12e       # show all hyperparams for this run
dora run -d -f 9357e12e hdemucs.channels=32   # new XP, new signature

Assessment

Why does Dora compute the signature as a delta from the base config rather than a hash of the full config? What breaks if you edit the base YAML directly?