Claude Skills by VectorSpaceLab

"Use TorchMetrics core APIs for import checks, metric lifecycle,

"Use TorchMetrics metrics that load pretrained models, feature

"Use TorchMetrics image quality, segmentation, object detection,

"Use TorchRL for TensorDict-first reinforcement-learning

"Collect rollouts, select collector topologies, configure replay

"Guide TorchRL code changes, tests, docs, CI labels, deprecations,

"Build, validate, transform, vectorize, and debug TorchRL

"Use TorchRL LLM post-training, VLA data/actions, services,

"Assemble TensorDict-native TorchRL modules, policies, recurrent

"Choose and wire TorchRL losses, value estimators, target updaters,

"Use torchtune for PyTorch-native LLM post-training, recipe/config

"Use torchtune's tune CLI, recipe registry, and OmegaConf config

"Build and validate torchtune dataset configs, message transforms,

"Use torchtune generation, Eleuther evaluation, and quantization

"Use torchtune model/tokenizer builders, PEFT modules, losses,

"Select and adapt torchtune post-training recipes without launching

"Use torchtune training utilities, checkpointers, precision,

"Use this skill when working with TorchVision models, weights,

"Use this sub-skill when constructing torchvision datasets,

"Select, construct, configure, inspect, and troubleshoot

"Use torchvision.ops for boxes, NMS, ROI pooling/alignment,

"Use when planning or auditing TorchVision reference

"Build torchvision transforms.v2 pipelines for images, videos,

"Use TotalSegmentator as an installed package for medical-image

"Plan TotalSegmentator nnU-Net retraining, evaluation, dataset

"Use TotalSegmentator auxiliary analysis CLIs for contrast phase,

"Discover TotalSegmentator tasks, classes, modalities, license

"Handle TotalSegmentator NIfTI and DICOM inputs, DICOM outputs,

"Consume TotalSegmentator masks, run reports, statistics,

"Install, inspect, and configure TotalSegmentator runtime

"Build reproducible TotalSegmentator CLI and Python API

"Route Towhee pipeline, operator, data utility, service, Triton,

"Use Towhee DataCollection, Entity, DataLoader, media wrappers,

"Operate Towhee ops wrappers, local operator registration, Hub

"Build, run, batch, debug, and smoke-check Towhee custom pipelines

"Construct Towhee API services and Triton-serving artifacts while

"Use Towhee's optional PyTorch Trainer, TrainingConfig YAML,

"Use Trafilatura for web text discovery, downloading, extraction,

"Use Trafilatura's command line for stdin, URL/list, and local

"Corpus-level quality, deduplication, structured output,

"Find, filter, retrieve, and politely crawl web pages with

"Use Trafilatura's Python APIs to extract text, metadata, comments,

"Route train-llm-from-scratch workflows for from-scratch PyTorch

"Guide JSON config editing, smoke configs, CLI override precedence,

"Prepare and validate flat pretraining, packed SFT, preference, and

"Guide GSM8K evaluation, answer verification, and checkpoint chat

"Guide architecture inspection and base-model pretraining from

"Guide the train-llm-from-scratch post-training alignment pipeline:

"Use Transfer Learning Library (TLLib) for PyTorch transfer

"Router for Transfer-Learning-Library domain adaptation workflows,