Skill UI

Collection of battle-tested PyTorch training recipes covering architecture selection, optimizers, LR scheduling, precision tips, and debugging to train or fine-tune LLMs, vision, diffusion, and biomedical models with high GPU throughput.

This skill guides the setup and execution of GPU-intensive Machine Learning workloads on Modal, a serverless cloud platform. It details a zero-config workflow for training, fine-tuning, and inference, eliminating the need for manual SSH setup or complex Docker orchestration. Key steps include detailed cost estimation, selecting the optimal GPU based on VRAM requirements, and generating reproducible Python launcher scripts for reliable ML experimentation.

This expert tool provides deep, systematic auditing of datasets to assess completeness, consistency, accuracy, and structural validity. It performs full data profiling, detects anomalies (outliers), analyzes missing data mechanisms (MCAR/MAR/MNAR), and calculates a Data Quality Score (DQS). Use it when validating data for model training, debugging pipelines, or establishing ongoing monitoring checks.

Interactive training for the GitHub Copilot CLI with guided lessons, quizzes, scenario challenges, and a reference covering slash commands, shortcuts, agents, MCP, and configuration; say “cliexpert” to begin.

Provides programmatic access to massive, versioned public single-cell and spatial transcriptomics datasets from CZ CELLxGENE. This skill allows users to efficiently query specific subsets of cell metadata, gene expression matrices, embeddings, and summary statistics across various organisms, tissues, and diseases. It is ideal for large-scale, memory-efficient biological research and model training without full dataset downloads.

This skill proactively detects and reports comprehensive system resources, including CPU cores, GPU types (NVIDIA, AMD, Apple Silicon), available memory, and disk space. It generates a detailed JSON report with strategic, context-aware recommendations for optimizing scientific computing workflows, such as recommending parallel processing libraries (Dask, joblib), out-of-core methods (Zarr), or appropriate GPU acceleration backends (Metal, ROCm). Use this before any large-scale data analysis, model training, or computationally intensive simulation to make informed architectural decisions.

PathML is a comprehensive Python framework designed for computational pathology workflows. It supports end-to-end analysis of Whole Slide Images (WSI), including loading 160+ proprietary formats, performing advanced stain normalization and tissue segmentation, constructing spatial cell graphs, and training deep learning models (e.g., HoVer-Net) for multiplex immunofluorescence and spatial proteomics. Ideal for large-scale biomedical image analysis.

A versatile, hardware-agnostic quantum computing framework designed for training quantum circuits and building hybrid quantum-classical models. It provides automatic differentiation, supporting advanced applications in Quantum Machine Learning, Quantum Chemistry (VQE), and variational algorithms, while ensuring portability across major quantum hardware backends (IBM, Google, etc.).

PufferLib is a high-performance, scalable RL framework designed for massive parallel training and vectorized simulation. Featuring PuffeRL (optimized PPO+LSTM) and the PufferEnv API, it supports multi-agent systems, custom environment creation, and achieves industry-leading speeds (up to millions of steps/second). Ideal for advanced research and large-scale AI experimentation.

A comprehensive framework designed to structure and automate PyTorch deep learning workflows. It simplifies boilerplate code by organizing models into LightningModules, automating training loops via the Trainer, and providing robust support for multi-GPU/TPU scaling (DDP, FSDP, DeepSpeed), data handling (DataModules), and best-in-class experiment logging (W&B, MLflow). Ideal for professional, scalable deep learning project development.

Stable Baselines3 is a robust, PyTorch-based library providing reliable implementations of state-of-the-art reinforcement learning algorithms (PPO, SAC, DQN, etc.). It is designed for standard RL experiments, quick prototyping, and developing complex single-agent systems. The library supports training agents, creating custom Gymnasium environments, and running vectorized simulations for accelerated learning.

This skill utilizes the TimesFM foundation model for zero-shot forecasting of any univariate time series. It predicts future values with calibrated probabilistic intervals without requiring custom model training. It includes a mandatory preflight system checker to ensure adequate RAM, GPU, and disk resources for safe, local inference across various data types (sales, sensor, energy).