Skill UI

This skill provides a managed cloud environment for training and fine-tuning advanced vision models. Users can build and customize models for object detection (e.g., DETR, YOLOS), image classification, and segmentation (SAM/SAM2). By leveraging Hugging Face Jobs, training is executed on powerful cloud GPUs without requiring local GPU setup, and all resulting models are automatically saved to the Hugging Face Hub.

This skill provides continuous, structured monitoring of deep learning model training metrics. It analyzes data from platforms like WandB and local logs to detect critical anomalies such as NaN values, divergence, plateaus, and sudden spikes. By providing a clear decision (CONTINUE, WAIT, or STOP) and detailed evidence, it ensures compute resources are only spent on viable training runs.

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 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.

Provides programmatic and standardized access to massive, versioned datasets of single-cell and spatial transcriptomics from CZ CELLxGENE. This skill allows users to query millions of cells, analyze gene expression, and access pre-calculated embeddings across various tissues, diseases, and species without downloading entire datasets. Essential for large-scale bioinformatics analysis, machine learning model training, and comprehensive cross-dataset comparison.

Histolab is a specialized Python library for digital pathology designed to process Whole Slide Images (WSI). It automates critical steps including sophisticated tissue detection, customizable tile extraction (random, grid, score-based), and stain normalization. It is essential for preparing high-quality, structured datasets for deep learning model training and comprehensive quantitative whole slide analysis.

Modal is a serverless cloud platform designed for executing Python-based AI/ML workloads. It provides on-demand access to powerful GPUs (e.g., H100, A100), scalable serverless functions, and custom containerization. Use cases include deploying ML models, running large-scale training/inference jobs, creating web APIs, and scheduling data pipelines, all defined entirely in code.

Molfeat is a comprehensive Python library designed for molecular featurization, providing over 100 pre-trained and hand-crafted featurizers. It unifies the process of converting chemical structures (SMILES strings or RDKit molecules) into robust numerical feature vectors. This capability is essential for advanced cheminformatics tasks, including Quantitative Structure-Activity Relationship (QSAR) modeling, virtual screening, similarity searching, and training deep learning models on molecular data.

PennyLane is a quantum computing library designed for building and training quantum circuits, particularly for quantum machine learning (QML) and quantum chemistry. It offers automatic differentiation capabilities for quantum circuits, making it ideal for developing hybrid quantum-classical models. It is hardware-agnostic, supporting backends from major providers like IBM, Google, and IonQ, and integrates seamlessly with classical ML frameworks such as PyTorch and JAX. Use cases include variational algorithms (VQE, QAOA), quantum neural networks, and molecular simulation.

PyTorch Lightning is an advanced deep learning framework built on PyTorch. It simplifies the entire ML workflow, allowing users to organize models into LightningModules, automate training loops using the Trainer, and manage complex data pipelines with DataModules. It supports advanced features like multi-GPU/TPU scaling (DDP, FSDP), integrated logging (W&B, MLflow), and custom callbacks, making large-scale neural network development efficient and reproducible.

Stable Baselines3 is a PyTorch-based library providing production-ready implementations of core reinforcement learning algorithms, including PPO, SAC, DQN, and TD3. It offers a scikit-learn-like API for easily training RL agents, developing custom Gymnasium environments, and managing complex training workflows. Ideal for standard RL research, rapid prototyping, and advanced model optimization.

Utilizes Google's TimesFM foundation model for zero-shot time series forecasting. It processes any univariate time series (e.g., sales, sensor data, weather) without requiring custom model training. The skill provides probabilistic forecasts with calibrated prediction intervals (quantiles) and includes a mandatory system checker to ensure sufficient RAM, GPU, and disk resources for safe local inference.