Hassan881/detection-images

Staging dataset for Berhan XAI (BurhanXAI) AI-generated / edited-media detection.

Field	Value
Training contract	v0.3.0 (firefly top-up applied; 6/6 M2 providers)
Held-out eval (M2)	v0.3.0-eval
Held-out eval (legacy)	v0.2.1-eval
Status	published
Classes	real, ai_generated, ai_edited
Seed	42

from datasets import load_dataset
ds = load_dataset("Hassan881/detection-images", name="v0.3.0")
ev = load_dataset("Hassan881/detection-images", name="v0.3.0-eval")  # test only

v0.3.0 training (post firefly top-up)

{
  "total": 14996,
  "counts_by_provider": {
    "stable_diffusion": 4000,
    "cifar10": 4000,
    "inswapper": 2000,
    "midjourney": 800,
    "dalle": 800,
    "gpt_image": 800,
    "flux": 800,
    "nano_banana": 800,
    "firefly": 996
  },
  "firefly_counts": {
    "train": 600,
    "val": 150,
    "test": 246
  },
  "all_six_ok": true
}

Firefly plan: train/val/test = {'train': 600, 'val': 150, 'test': 246} (stock 996 unique; test short by 4 vs ideal 250).

v0.3.0-eval (M2 held-out)

Sha256-disjoint from amended v0.3.0 (train+val+test) and v0.2.0 (train+val). Target ~200 per M2 provider.

{
  "midjourney": 200,
  "dalle": 200,
  "gpt_image": 200,
  "flux": 200,
  "nano_banana": 200
}

Empty/short strata: ['firefly: 0/200 (synthbuster 1000 unique stock fully used by v0.3.0 top-up 600/150/246)']

Do not delete prior Hub revisions/configs. Trainer loads name="v0.3.0".