ai-ml

Plugin ai-ml from marketplace ai-skills in CHENyiru3/AI-Skills-Collections. Commands read from the repo's manifest — 3 days ago.

Skills (14)

• jepa-pytorch-development

  Design, implement, adapt, or debug JEPA representation learners and latent world models in PyTorch. Use for masked EMA-teacher JEPA, V-JEPA dense/deep supervision, teacher-free LeJEPA/SIGReg, or action-conditioned latent prediction, including token and mask contracts, gradient flow, collapse diagnos

• pytorch

  Implement or debug PyTorch tensor operations, nn.Module models, autograd, custom training loops, checkpoints, and device or precision issues. Use for concrete PyTorch development; prefer a project-specific model skill when its contracts apply.

• bitsandbytes

  Configure or debug bitsandbytes 4-bit/8-bit model loading, quantization settings, supported hardware, and PEFT/QLoRA preparation. Use for an actual bitsandbytes integration rather than generic model optimization.

• browser-use

  Configure or debug the browser-use Python agent, its model provider, and Chromium/CDP browser session. Use when browser-use is explicitly requested or already used in the project; not for general browser tasks with an available browser tool.

• huggingface-hub

  Download, inspect, version, or publish model/dataset/Space repositories using huggingface_hub or the hf CLI. Use for Hub repository operations, access, cache handling, and model-card metadata.

• peft

  Configure, train, load, switch, merge, or debug Hugging Face PEFT adapters such as LoRA, QLoRA, prompt tuning, and prefix tuning. Use when an actual PEFT adapter workflow is requested.

• transformers

  Implement or debug Hugging Face Transformers model/tokenizer loading, task heads, generation, feature extraction, or Trainer fine-tuning. Use in a Transformers codebase or when that library is requested.

• trl

  Implement or debug TRL supervised fine-tuning, preference optimization, reward modeling, or reinforcement-learning trainers, including their dataset schemas and evaluation. Use for a specific TRL workflow.

• latent-identifiability

  Develop, prove, or audit identifiability claims for latent-variable and representation-learning models, including VAEs, nonlinear ICA, and spatial or multi-view biological models. Use to turn a model idea into a quantified theorem, identify the remaining symmetries, construct a lemma-based proof, or

• vi-methodology-development

  Formulate, implement, or review a variational-inference or VAE research method, from generative assumptions and ELBO to posterior queries and validation. Use for method development or scvi-tools model integration; not routine use of an existing model.

• accelerate

  Configure or debug Hugging Face Accelerate launchers, Accelerator training loops, mixed precision, accumulation, distributed metrics, and checkpoint resume. Use when the project uses Accelerate.

• datasets

  Load, transform, split, stream, cache, or export data with Hugging Face Datasets. Use for Dataset/DatasetDict/IterableDataset workflows and their schema, batching, or reproducibility issues.

• deepspeed

  Configure or debug DeepSpeed engines, ZeRO stages, offloading, mixed precision, checkpointing, or a supported Transformers/Accelerate integration. Use for an actual DeepSpeed workload.

• pytorch-lightning

  Implement, migrate, or debug LightningModule, LightningDataModule, Trainer configuration, callbacks, logging, checkpoint resume, and distributed training in a PyTorch Lightning project.