Write benchmark reproduction requirement documents from a benchmark name plus an optional target method. Use when the user wants to reproduce, benchmark, evaluate, or plan work on a public benchmark and needs a requirement doc, latest literature/results summary, a reproduced method target, comparison baselines, reference reproduction methods, or an algorithm choice between SFT, DPO, and GRPO. If the user does not provide a method, first research the latest academic methods tested on that benchmark, choose a reproduction target, and then pick a practical execution baseline that can be pushed toward that target effect.
Reconcile same-named shared helpers across experiments, scaffold templates, profiles, and repo docs without silently changing the contract under the same name. Use when repo-wide helper names drift, when a helper is promoted into `scaffolds/single_file_experiment/naming.md`, or when experiment/template/ profile/skill updates must stay synchronized. Core behavior: scan repeated helper names, classify shared-contract drift vs task-specific differences, choose a canonical implementation, patch live code first, then sync templates/docs/skills, and verify the aligned helpers with AST/tests/hash checks.
MinT SDK API reference and production training patterns. Use when writing or modifying training scripts that call `import mint`: correct API signatures, loss functions, data types, async concurrency patterns, checkpoint management, and debugging. Covers: SFT (cross_entropy), RL/GRPO (Tinker-aligned `importance_sampling`, with `ppo` called out separately when needed), custom loss (forward_backward_custom for DPO), async concurrent sampling with sliding-window concurrency, session recovery, and framework migration (verl/TRL → MinT).
Upgrade or strengthen a MinT cookbook experiment into a research-ready harness under `experiments/`. Use when an experiment needs infrastructure beyond what scaffolds/profiles provide: resume and recovery support, checkpoint registries, richer run outputs, **research-grade structured logs** (SFT row cadence, merged train+eval rows, run-scoped JSONL), throughput-aware train design, or clearer benchmark and provenance contracts. Helps authors **design and document** research-level operability and logging (README result sections, `run.json` provenance, SFT/DPO/RL checklists) and align SFT with tinker-cookbook supervised patterns without importing that repo into experiments. This skill does not define baseline artifact naming or an alternate scaffold. It upgrades experiments that already follow the repo's canonical `scaffolds/` contract.
Create a new self-contained MinT cookbook experiment under `experiments/`. Covers the full eval-first lifecycle: scaffold the experiment directory, establish the eval baseline with the scaffold-owned eval artifact set, and add basic SFT, DPO, or GRPO training by following `scaffolds/profiles/`. This skill materializes experiments from the repo-level `scaffolds/` directory. It does not own a separate template tree.
Create publication-ready experiment figures for top-tier conference and journal papers. Use when Codex needs to draw or redraw training curves, evaluation curves, benchmark comparisons, ablation bars, task breakdowns, delta plots, or compact multi-panel figures from experiment artifacts, JSONL logs, CSV tables, or paper result tables. Prefer this skill when the figure must balance paper-grade rigor, clean statistical semantics, vector export, and fast visual scanning.
Define and bootstrap the repo-level harness for this MinT cookbook experiment monorepo. Use when initializing the repository, retrofitting the root-level conventions, or clarifying how independent experiment subdirectories should be structured, run, benchmarked, and maintained. Core behavior: create a minimal root harness, make each `experiments/<name>/` directory self-contained and easy for a developer or agent to enter, and make `scaffolds/` the source of truth for experiment templates.