---
license: other
license_name: mixed-upstream
task_categories:
  - video-classification
  - audio-classification
tags:
  - deepfake-detection
  - out-of-distribution
extra_gated_prompt: >-
  This dataset aggregates media from third-party sources under differing
  terms, and contains recordings of real people. Accepting here does not
  relieve you of complying with each upstream source's own licence.
extra_gated_fields:
  I will use this dataset for non-commercial research only: checkbox
  I will comply with each upstream source's own licence: checkbox
---

# DeepSafe Wild Test Set

A held-out set used to check
[DeepSafe](https://github.com/deepsafehq/deepsafe-bench) against media it was
never tuned on.

## What is in it

| Source | Files | Type |
|---|---|---|
| DeepAction v1 | 1,966 | AI-generated video across several generators |
| Gary Stafford audio | 1,011 | synthetic and real speech |
| Defactify | 17 | mixed |

## Why it is separate

The main [evaluation dataset](https://huggingface.co/datasets/deepsafe/evaluation-dataset)
trained the ensemble meta-learners. Measuring on it alone would report
in-distribution performance and overstate what the detectors do in the wild.
This set exists to be the part nothing was fitted to.

The gap between the two is the finding: across 411 generators the ensemble
catches 66% of fakes overall, and only 7% of Sora video. Full results in
[BENCHMARK.md](https://github.com/deepsafehq/deepsafe-bench/blob/main/BENCHMARK.md).

## Takedown

If you hold rights to any included material, open an issue at
https://github.com/deepsafehq/deepsafe-bench. Removal within 48 hours.