quick-research-campaign

The hardest part of ML parameter tuning is rarely “finding a good config.” It is usually the aftermath: twenty trials later you realize gate_type never tested one of its values, rope_base is perfectly locked to it, and the ablation table simply does not exist. quick-research-campaign turns that mess into a pipeline—define the target, pick a repo, align the baseline, make the minimal change, sweep at scale, pick the winner, attribute via the ingredient list, and ship the write-up—with a tool backing each step instead of relying on chat-log archaeology.

It ships three tools and one loadable skill. research_sweep_plan expands the search space into a bounded, de-confounded trial matrix: full cartesian product when the budget allows, uniform sampling rather than fixed-step truncation when it does not, plus explicit reports on which values were never tested and which parameter pairs are confounded. research_sweep_record logs every run as ok, failed, or pruned with metrics—failures and pruned runs count, because an unbounded search space cannot support an experimental section. research_sweep_report ranks by a single metric, prints the gap against the pinned baseline and SOTA target, flags whether the top ranks sit inside noise, computes each parameter’s marginal effect, runs a confound check, and emits a paper material checklist where every unchecked item is exactly what your methods section still owes.

Built for researchers and engineers running systematic tuning campaigns who need a reproducible ablation table and a defensible “why this config” narrative. It deliberately does not train, schedule, or build UIs: the ledger is per-session process state, and the agent plus shell do the actual running. You handle the hands-on work; it handles making sure you can explain the result.