run-train
๐ฏSkillfrom lllllllama/rigorpilot-skills
|A RigorPilot trusted-lane skill for conservatively starting or verifying deep learning training runs, including resume handling, bounded monitoring, and generating durable training records.
Overview
run-train is a trusted-lane skill in the RigorPilot suite designed for conservatively starting, resuming, or verifying deep learning training runs. It operates under the RigorPilot principle of preserving scientific meaning and comparability, ensuring that training sessions produce durable, auditable records without introducing unverified changes to the experimental setup.
Key Features
- Conservative training execution: Starts or resumes training runs using only documented commands and configurations from the repository, avoiding any silent modifications to hyperparameters or training logic.
- Resume handling: Detects and manages checkpoint-based resume scenarios, verifying that resumed runs maintain consistency with prior training state.
- Bounded monitoring: Tracks training progress within predefined bounds, surfacing anomalies and blockers without overriding researcher decisions.
- Durable training records: Generates structured output bundles in
train_outputs/that include logs, metrics, and configuration snapshots for reproducibility audits. - Two-lane routing: Follows the RigorPilot lane model where training runs default to the trusted lane, requiring explicit researcher authorization before any exploratory modifications.
Who is this for?
- Deep learning researchers who use AI coding agents and want guardrails that prevent silent scientific drift during training.
- Teams reproducing published results who need verifiable, auditable training records that match documented procedures.
- ML engineers managing long-running training pipelines who need safe resume and monitoring capabilities without risking comparability.
Same repository
lllllllama/rigorpilot-skills(14 items)
Installation
npx vibeindex add lllllllama/rigorpilot-skills --skill run-trainnpx skills add lllllllama/rigorpilot-skills --skill run-train~/.claude/skills/run-train/SKILL.mdSKILL.md
More from this repository10
The primary explore-lane orchestrator in the RigorPilot suite that enables meaningful and potentially novel exploration on top of current deep learning research, with idea gating, bounded experiments, and candidate ranking.
The primary end-to-end orchestrator in the RigorPilot suite that guides AI agents through README-first reproduction of deep learning repositories, ensuring scientific rigor, comparability, and auditable outputs.
A RigorPilot explore-lane skill for candidate implementation, code transplanting, and stitching on isolated branches when the researcher has explicitly authorized exploratory work on top of current research.
A helper skill in the RigorPilot suite that resolves gaps between a deep learning repository's README and its associated research paper, helping agents understand the paper context needed for faithful reproduction.
A RigorPilot trusted-lane skill for research-safe debugging of deep learning experiments that analyzes failures first and applies patches only after explicit approval, preserving scientific comparability.
A RigorPilot trusted-lane skill that performs read-only project analysis on deep learning repositories, including model mapping, structure inspection, and risk surfacing, without making any edits.
A RigorPilot helper skill that scans deep learning repositories and extracts README commands to build an initial reproduction plan, typically invoked by the main reproduction orchestrator.
A RigorPilot trusted-lane skill for conservatively setting up deep learning environments, including datasets, pretrained weights, checkpoints, and cache assumptions required for research reproduction.
A RigorPilot explore-lane skill for running small-subset probes, short-cycle trials, and ranked exploratory experiments when the researcher explicitly authorizes candidate-only exploration.
A RigorPilot trusted-lane skill for conservatively running documented evaluation, inference, smoke tests, and sanity checks in deep learning repositories without introducing unaudited changes.