Claude Skills by VectorSpaceLab

"Export trained YOLOX checkpoints to ONNX or TorchScript and reason

"Use YOLOX for PyTorch inference, demo CLI/API setup, model

"Design and validate YOLOX training, data, Exp, caching, logging,

"Operate and maintain Yuxi, a Docker Compose managed FastAPI,

"Implement/debug Yuxi agent runtime, APIs, tools, skills, MCP,

"Operate and extend yuxi-cli plus the external API/SSE integration

"Boot, configure, and operate Yuxi Docker, Lite, production, and

"Operate Yuxi knowledge-base, retrieval, document parsing, OCR,

"Maintain the Yuxi monorepo: Docker Compose workflows,

"Route ZenML repository and package tasks for MLOps pipelines, CLI

"Use ZenML from the CLI and Python Client for repository

"Build, adapt, validate, and troubleshoot ZenML pipeline

"Maintain the ZenML repository safely with AGENTS guidance,

"Build, configure, run, schedule, debug, and validate ZenML step

"Work on ZenML FastAPI server routers, auth/RBAC, domain models,

"Register, configure, and extend ZenML stacks, stack components,

"Guides Zero123Plus generation, ControlNet variants, demo launch,

"Guides Zero123Plus demo, serving, and Cog deployment workflows,

"Guides Zero123Plus single-image to multi-view generation,

"Use ZhuSuan for Bayesian networks, variational inference,

"Use ZhuSuan for HMC, SGLD, PSGLD, SGHMC, SGNHT, annealed

"Use ZhuSuan's distributions, BayesianNet, MetaBayesianNet, and

"Use ZhuSuan for ELBO, IWAE, inclusive KL, importance-sampling

"Operate zi2zi legacy Chinese/Japanese/Korean font style transfer

"Prepare zi2zi paired glyph images and pickle streams from

"Run zi2zi checkpoint inference, style interpolation, transition

"Train, fine-tune, and understand zi2zi's TensorFlow 1.x

Recover solutions after a focused attempt lacks a verified full pass or maximum score. MUST invoke and read this skill (1) after 2 non-full submissions, even if high-scoring; (2) after the first non-full candidate for any interactive or scored heuristic/optimization task, including AHC; and (3) before any non-full final delivery. Skip only after verified full pass/full score. For (2), route to an independent local interactor/checker/scorer only on concrete evaluator, legality, scoring, protoc...

Design, implement, debug, and validate independent C++ checkers, scorers, interactors, simulators, and local evaluation workflows for algorithmic and competitive-programming tasks. Use when concrete evidence makes evaluator reconstruction diagnostic: an official evaluator is absent or doubtful and blocks a decision, zero/invalid/WA or impossible feedback is unexplained, local and remote behavior disagrees, output legality or score must be reconstructed, an interactive protocol fails, or Plate...

Diagnose and harden contest or bounded algorithmic solver implementations for compiler and toolchain mismatches, crashes, undefined behavior, overflow, memory-layout failures, TLE or MLE, buffering, serialization, randomness, data-structure invariants, and unsafe deadline handling. Use as an internal algorithmic-problem-solving recovery route after evidence implicates the implementation or target environment rather than the mathematical model, evaluator contract, or search-method design.

Solve or debug information-acquisition tasks where queries or evaluator feedback must reduce uncertainty until a hidden state, answer, or contract is sufficiently determined. Use for interactive problems, output-only tasks with a checker, and custom-judge workflows when the central difficulty is the protocol contract, query design, hypothesis elimination, query budget, transcript behavior, noise, or adversarial feedback. Do not use for offline output-only or scorer-only optimization whose fee...

Diagnose and repair a static algorithmic problem model, state, reduction, recurrence, invariant, or proof, then decide whether an exact, constructive, exact-hybrid, heuristic, or scored-hybrid route fits the real bounds. Use as an internal algorithmic-problem-solving recovery route when brute force or a proof premise contradicts the attempted solution, or when the model is trusted but the justified solution class remains unclear. Do not use for hidden-state query design or reward-bearing onli...

Escape a root-verified measured plateau or severe score gap in a scored algorithmic, heuristic, or hybrid solver through independent review, method research, and an evidence-gated structural prototype. Use only when the root router admits a verified legal champion. For an interactive or scored heuristic/optimization task, including AHC, first implement and validate its mandatory independent local interactor/checker/scorer under checker-and-local-evaluation, delegating missing implementation t...

Design, debug, and improve decision policies that must earn immediate or future reward under uncertainty, including reactive or online control, repeated-round learning, exploration, and risk-aware action selection. Use when actions occur in a live sequential process, their feedback changes later choices in the same run or environment, and termination is a horizon, task completion, or decision budget. Do not use for offline output-only AHC or scorer-only optimization across development runs; u...

Guide C++ competitive-programming judging with the bundled testlib.h. Use when the user asks to author special or scored checkers, strict validators, deterministic generators, basic interactors, or a minimal local solution-checker workflow.

Design the smallest falsifying test loop for an algorithmic solver, including brute oracles, differential and metamorphic tests, champion/challenger comparisons, paired seeds, holdouts, resource probes, and release gates. Use as an internal algorithmic-problem-solving recovery route when the smallest counterexample is unclear, comparisons are noisy, or a correction needs independent evidence before promotion. Do not use to implement a doubtful checker or to diagnose a known compiler or runtim...

Run bounded soft-mode AdaLoRA recovery experiments with source-boundary logs, mechanism checks, and validation artifacts.

Schedule AdaLoRA global rank budgets with warmup, cubic decay, mask intervals, and final fixed allocation.

Compute AdaLoRA sensitivity-uncertainty triplet scores and globally mask singular values to a target rank budget.

Implement AdaLoRA SVD-style low-rank linear updates with frozen base weights and active-rank scaling.

Run a bounded CoFi proxy experiment that exercises masks, sparsity regularization, distillation, pruning summaries, and validation logs.

Compute CoFi target-sparsity warmup, expected sparsity, and Lagrangian penalty/update diagnostics for pruning runs.

Build dynamic teacher-student layer alignments and representation distillation losses for pruned transformer students.

Compose CoFi coarse and fine pruning masks and summarize active transformer structures for reduced recovery experiments.

Evaluate LoRA parameter efficiency, rank constraints, merge checks, and proxy recovery metrics.

Create and validate low-rank LoRA adapters for linear layers with merge-equivalent inference.

Run a bounded LoRA-only optimization loop and log mechanism-faithful training evidence.

Freeze pretrained parameters and expose only LoRA task parameters for training or checkpointing.

Aggregate LoRAPrune element importances into structured groups and create deterministic channel or head pruning masks.

Compute LoRAPrune's LoRA-gradient-only Taylor importance estimate for structured pruning without frozen base-weight gradients.