# Create FiftyOne Inference Datasets

Computer vision teams need repeatable ways to load media, labels, and predictions into FiftyOne. This skill guides Claude, Codex, and Claude Code through dataset creation, import validation, and model inference.

## Install

```bash
npx skillstore add adonaivera/fiftyone-dataset-inference
```

## Metadata

- Status: approved
- Slug: adonaivera-fiftyone-dataset-inference
- Skillstore revision: r1
- Version status: missing
- Tree hash: aa68a23eb349bf828de1be2d912c8319af35f7c845524ea63011c03a9ad51fad
- Author: AdonaiVera
- GitHub username: AdonaiVera
- License: MIT
- Repository: https://github.com/AdonaiVera/fiftyone-skills/tree/main/dataset-inference/skills/fiftyone-dataset-inference
- Ref: 6eec4a59f63ff83440a84c18d1aef79743bc714e
- Supported tools: Claude, Codex, Claude Code
- Audit status: complete
- Agent install advisory: allowed
- Manual install advisory: allowed
- Artifact signature: available
- Audit attestation: unavailable
- Human verification: not\_verified
- Risk factors: external\_commands, network
- Quality score: 69
- Public page: https://skillstore.pages.dev/skills/adonaivera-fiftyone-dataset-inference
- Manifest: https://skillstore.pages.dev/api/skills/adonaivera-fiftyone-dataset-inference/manifest

## Capabilities

- Scans local directories to identify supported media and label files.
- Creates persistent FiftyOne datasets through MCP operators.
- Imports media-only datasets and labeled COCO, YOLO, VOC, CVAT, KITTI, and TF formats.
- Runs built-in FiftyOne model zoo inference for detection, classification, segmentation, and embeddings.
- Launches the FiftyOne App and filters views to inspect prediction fields.
- Validates imported sample counts against the source directory.

## Use Cases

- Prepare Image Review Sets: Load unlabeled image folders into FiftyOne and generate detection predictions for visual review.
- Import Labeled Benchmarks: Create datasets from COCO, YOLO, or VOC labels while keeping ground truth separate from predictions.
- Validate Client Data Drops: Inspect folder contents, import samples, confirm counts, and open a FiftyOne view for quality checks.

## Prompt Templates

### Import Media Folder

```
Create a FiftyOne dataset from this local media folder. Inspect the folder first, propose a dataset name, and ask before importing.
```

### Import COCO Labels

```
Import my COCO image dataset into FiftyOne. Use a separate ground_truth field and validate the sample count after import.
```

### Run Fast Detection

```
After the dataset is imported, run a fast built-in detection model and save predictions in a predictions field.
```

### Build Full Review Workflow

```
Build an end-to-end FiftyOne workflow for this dataset. Detect the format, import labels, run a suitable model, launch the app, and summarize validation results.
```

## Limitations

- Requires a running FiftyOne MCP server and the referenced Voxel51 plugins.
- Only supports dataset types and model names available in the local FiftyOne environment.
- Some models require extra packages, hardware resources, or longer processing time.
- The skill does not evaluate model accuracy or tune model thresholds automatically.

## Best Practices

- Confirm dataset name, label fields, and source paths with the user before import.
- Keep ground truth labels and model predictions in separate fields.
- Start with smaller model variants before running slower or heavier models.

## Anti Patterns

- Do not overwrite existing datasets without explicit user approval.
- Do not run dependency install commands without reviewing the exact command first.
- Do not mix imported labels and predictions in the same field.

## Security Audit

- Audited at: 2026-07-05T01:27:13.432\+00:00
- Summary: The static findings are mostly false positives caused by Markdown fences, inline code, and documentation links. The skill is focused on FiftyOne dataset import and inference, but it includes behavior-level risks around Bash directory scans and dependency installation commands returned by tooling.

## Stats

- Views: 327
- Downloads: 10
- Favorites: 0
- Popularity score: 0