voxel51/fiftyone-skills

Write Python code following FiftyOne's official conventions. Use when contributing to FiftyOne, developing plugins, or writing code that integrates with FiftyOne's codebase. Covers module structure, import organization, Google-style docstrings, lazy imports, guard patterns, and error handling.

Creates Jupyter notebooks for FiftyOne workflows including getting-started guides, tutorials, recipes, and full ML pipelines. Use when users want to create a notebook, write a tutorial, build a demo, document a workflow, or generate a FiftyOne walkthrough covering data loading, exploration, inference, evaluation, and export. Builds notebooks cell-by-cell using NotebookEdit.

End-to-end dataset curation for any FiftyOne dataset: inspect schema and media quality (blur, brightness, corruption, resolution), audit annotation quality (mistakenness, hardness, duplicate labels, IoU overlaps), analyze class distributions and imbalances, explore embedding space and data gaps, find near/exact duplicates, create curated subsets using semantic search and brain ops, build train/val/test splits, and answer natural language questions about your data. Delegates to fiftyone-find-duplicates and fiftyone-embeddings-visualization. Requires FiftyOne MCP server with @voxel51/brain and @voxel51/utils plugins.

Export FiftyOne datasets to standard formats (COCO, YOLO, VOC, CVAT, CSV, etc.) and Hugging Face Hub. Use when converting datasets, exporting for training, creating archives, sharing data in specific formats, or publishing datasets to Hugging Face Hub.

Universal dataset import for FiftyOne supporting all media types (images, videos, point clouds, 3D scenes), all label formats (COCO, YOLO, VOC, CVAT, KITTI, etc.), multimodal grouped datasets, and Hugging Face Hub datasets. Use when users want to import any dataset regardless of format or source, load datasets from Hugging Face, handle autonomous driving data with multiple cameras and LiDAR, or create grouped datasets from multimodal data.

Run model inference on FiftyOne datasets using Zoo models or custom models. Use when applying ML models for detection, classification, segmentation, or embeddings on existing datasets.

Develop custom FiftyOne plugins (operators and panels) from scratch. Use when users want to create, build, or develop a new plugin for FiftyOne App, extend FiftyOne with custom operators, create interactive panels, add new functionality to FiftyOne, or integrate external APIs/services into FiftyOne.

Visualize datasets in 2D using embeddings with UMAP or t-SNE dimensionality reduction. Use when users want to explore dataset structure, find clusters in images, identify outliers, color samples by class or metadata, or understand data distribution.

Evaluates FiftyOne plugins for quality, security, and agent-readiness. Use when reviewing a plugin before installation, auditing an existing plugin, validating a plugin you just built, or assessing community plugins for safety. Checks manifest integrity, security risks (filesystem access, network calls, command execution, data exfiltration), schema quality, risk classification, code conventions, and agent discoverability. Produces a structured report with scores and actionable recommendations.

Find duplicate or near-duplicate images in FiftyOne datasets using brain similarity computation. Use when users want to deduplicate datasets, find similar images, cluster visually similar content, or remove redundant samples. Requires FiftyOne MCP server with @voxel51/brain plugin installed.

Generate a Data Lens connector from an external database schema. Use when users want to connect an external data source (PostgreSQL, BigQuery, Databricks, MySQL, SQLite, etc.) to FiftyOne Data Lens, create a DataLensOperator from a CREATE TABLE statement or column list, or build a plugin that lets users browse and import data from their database through the FiftyOne App.

Triage FiftyOne GitHub issues by validating status, categorizing resolution, and generating standardized responses. Use when reviewing issues to determine if already fixed, won't fix, not reproducible, no longer relevant, or still valid. Includes investigation workflow and response templates.

Evaluate model predictions against ground truth using COCO, Open Images, or custom protocols. Use when computing mAP, precision, recall, confusion matrices, or analyzing TP/FP/FN examples for detection, classification, segmentation, or regression tasks.

Diagnose and fix common FiftyOne issues automatically. Use when a dataset disappeared, the App won't open, changes aren't saving, MongoDB errors occur, video codecs fail, notebook connectivity breaks, operators are missing, or any recurring FiftyOne pain point needs solving. Includes a NEVER-DO safety section to prevent accidental data loss or direct MongoDB manipulation.

Build FiftyOne UIs using VOODO (@voxel51/voodo), the official React component library. Use when building plugin panels, creating interactive UIs, styling FiftyOne applications, or needing React components with proper design tokens. Fetches complete component API reference dynamically.