README.md

metadata

license: other
private: true
task_categories:
  - image-classification
language:
  - en
tags:
  - deepfake
  - df40
  - real-vs-fake

DF40 GretchenDefender Dataset

Private compressed dataset containing DF40 train/val JSONL splits and referenced images.

• Total JSONL rows: 2,536,291
• Unique images: 1,543,531
• Missing images at prepare time: 0

Splits

File	Rows	Labels
train_DF40_EFS.jsonl	410,000	Real: 229,930, Fake: 180,070
train_DF40_FE.jsonl	28,754	Fake: 28,754
train_DF40_FR.jsonl	545,522	Real: 275,916, Fake: 269,606
train_DF40_FS.jsonl	413,599	Real: 206,937, Fake: 206,662
val_DF40_EFS.jsonl	81,900	Real: 44,790, Fake: 37,110
val_DF40_EFS_cdf.jsonl	323,793	Real: 56,200, Fake: 267,593
val_DF40_FR.jsonl	107,056	Real: 53,748, Fake: 53,308
val_DF40_FR_cdf.jsonl	309,872	Real: 67,440, Fake: 242,432
val_DF40_FS.jsonl	101,789	Real: 40,311, Fake: 61,478
val_DF40_FS_cdf.jsonl	214,006	Real: 50,580, Fake: 163,426

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

# JSONL
mkdir -p jsonl && gzip -dc compressed/jsonl/train_DF40_EFS.jsonl.gz > jsonl/train_DF40_EFS.jsonl

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

JSONL format

{
  "messages": [
    {
      "role": "user",
      "content": "<image>Is this image real or fake?"
    },
    {
      "role": "assistant",
      "content": "Real"
    }
  ],
  "images": [
    "images/processed/FF++/original_sequences/youtube/c23/frames/071/277.png"
  ]
}