Skill UI

Automated guidance for managing model checkpoints in ML training workflows, offering best practices, sample code, and validation to keep experiments stable and reproducible.

A comprehensive guide to managing training data and model artifacts securely on cloud GPU instances (e.g., Vast.ai). It covers best practices for data transfer (SCP, compressed, cloud staging), implementing AES-256 encryption, managing checkpoints to S3, and performing secure cleanup before instance destruction, ensuring compliance and data integrity.

A comprehensive guide detailing best practices for ensuring the reproducibility of computational claims in Large Language Model (LLM) papers. It outlines a layered approach (Exact, Budgeted, Drifting) for verifying results, providing protocols for handling open-weight checkpoints, proprietary API models, and documenting compute costs and failure modes honestly.

This comprehensive guide outlines best practices for packaging and releasing scientific artifacts—including code, training configurations, demonstration data, and policy checkpoints—for robotics research papers. It details how to prepare materials for anonymous reviewer assessment and for durable public release, ensuring the work is fully reproducible by the academic community.

A comprehensive guide on preparing and structuring evidence packages for recommender system research, especially for academic submissions. It details how to ensure that all ranking claims, off-policy estimates, and evaluation metrics are end-to-end reproducible by providing artifacts like datasets, split scripts, model checkpoints, and propensity logs in an anonymous, verifiable repository.

Work with Outpost Bio's open microbiome foundation models: embed taxonomic profiles, fine-tune Waypoint checkpoints, benchmark on Compass, pretrain on Atlas, and convert MetaPhlAn, Kraken2, QIIME 2, or MGnify tables via the waypoint CLI.

This skill enables independent auditing of AI agent, gateway, or automation workflow runs. It reconstructs events from traces, checkpoints, and tool calls to verify success claims without trusting self-reports. Users establish contracts, build claim ledgers, normalize evidence, rank witnesses, and assign verdicts like proven or contradicted. It ensures integrity in multi-step automation and separates autonomous success from human-assisted outcomes.

Condense a long agent conversation at a verified phase boundary by writing a durable checkpoint that preserves constraints, evidence, decisions, artifacts and the next action, so a fresh context can resume the same task without repeating work or losing state.

Plans first, then orchestrates dynamic multi-agent workflows with adversarial verification via a local odw daemon or native fallback. Supports 16-100 parallel agents, sandboxed scripts, SQLite checkpoints, and crash-resume for tasks spanning many files or items.