Skill UI

bdistill transforms AI interactions into a structured, queryable knowledge base. It extracts, structures, and quality-scores domain knowledge from both closed-source AI models (in-session) and local open-source models (via Ollama). Users can build comprehensive reference datasets, Q&A pairs, or generate structured training data for traditional machine learning, ensuring validated and exportable domain expertise.

Trackio is an essential library for logging, visualizing, and analyzing machine learning training experiments. It provides multi-interface support: a Python API for real-time metric logging and alert firing during training; a Command Line Interface (CLI) for retrieving historical metrics and alerts; and synchronization capabilities to Hugging Face Spaces for persistent dashboards. This tool is crucial for debugging, diagnosing training issues (like loss spikes), and automating ML workflow iteration.

This skill automates the process of renting, managing, and terminating cloud GPU instances on vast.ai. It analyzes your AI/ML task—determining VRAM, required GPU count, and estimated runtime—and then searches for the most cost-effective and reliable GPU offers, presenting optimized options for your training workflow.

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.

Utilize Modal, a serverless GPU cloud, for running complex machine learning workloads, including training, fine-tuning, and inference. This tool provides a zero-config environment, eliminating the need for SSH setup or Docker management. It features automatic scale-to-zero billing and local-first execution, making high-powered GPU compute accessible for debugging and small-scale experiments.

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 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.

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).

This guide details how to set up and execute large-scale, distributed GPU training jobs on the CoreWeave platform. It covers both single-node multi-GPU setups and multi-node training utilizing PyTorch DDP and Kubernetes resources (A100/H100). This workflow is ideal for fine-tuning large language models (LLMs) or running complex deep learning models that require high-performance, shared-storage compute clusters.

A comprehensive tool for creating high-quality training datasets for custom object detection and segmentation models. It supports multiple advanced annotation methods including BBox drawing, SAM2 pixel-perfect segmentation, and DINOv3 visual grounding. Features include AI auto-detection, object tracking, and standardized COCO format export, streamlining the entire data pipeline from capture to training.