experiment-bridge

Automating experiment planning from markdown to GPU deployment requires parsing EXPERIMENT_PLAN.md to generate executable training code, optionally reviewing it, and queueing GPU runs. This Skill bridges that gap by converting milestones and hyperparameters into deployed runs with initial metrics.

Yes, you can run a sanity check before main training runs. This Skill optionally executes a small sanity-stage experiment first and auto-debugs retries if it fails, ensuring baseline logic and evaluation ground-truth are validated before committing to full GPU deployment.

Workflow automation supports parallel GPU deployment for ablation studies. Experiments defined in the plan are deployed in parallel or via an orchestrated queue, updating tracker logs with parseable JSON/CSV outputs for iterative review across multiple runs.

To collect evaluation metrics for research iteration automatically, the Skill deploys runs to GPU infrastructure and parses initial outcomes into JSON/CSV formats. It updates the experiment tracker with these metrics for downstream automated review loops.

You need FINAL_PROPOSAL.md to use the experiment bridge workflow. The Skill reads FINAL_PROPOSAL.md for method context and EXPERIMENT_PLAN.md to determine run order, datasets, and required hyperparameters before implementing and deploying training code.

The best way to prevent logic mistakes in automated training runs is applying cross-model code review guardrails before deployment. This Skill optionally performs a code-review step to reduce evaluation-ground-truth and logic errors prior to executing GPU runs.