README.md

---
license: other
private: true
task_categories:
  - image-classification
language:
  - en
tags:
  - deepfake
  - ff++
  - faceforensics
  - real-vs-fake
---

# FF++ GretchenDefender Dataset

Private compressed dataset containing FF++ train/val JSONL splits and referenced images.

- Total JSONL rows: 137,238
- Unique images: 137,238
- Missing images at prepare time: 0

## Splits

| File | Rows | Labels |
|---|---:|---|
| `train_FF++.jsonl` | 114,884 | Real: 22,993, Fake: 91,891 |
| `val_FF++_val.jsonl` | 22,354 | Real: 4,477, Fake: 17,877 |

## Layout

- `compressed/jsonl/*.jsonl.gz` - gzipped JSONL splits
- `compressed/images/images_shard_XXXX.tar.gz` - image shards
- `compressed/manifest.jsonl.gz` - absolute to repo path mapping
- `shards_index.json` - shard metadata

## Extract

```bash
# JSONL
mkdir -p jsonl && gzip -dc compressed/jsonl/train_FF++.jsonl.gz > jsonl/train_FF++.jsonl

# Images
for shard in compressed/images/images_shard_*.tar.gz; do
  tar -xzf "$shard"
done
```

## JSONL format

```json
{
  "messages": [
    {
      "role": "user",
      "content": "<image>Is this image real or fake?"
    },
    {
      "role": "assistant",
      "content": "Real"
    }
  ],
  "images": [
    "images/processed/FF++/original_sequences/youtube/c23/frames/004/000.png"
  ]
}
```