{
  "version": "v5.1",
  "description": "AI Content Detection meta-classifier weights (stacking ensemble)",
  "modalities": [
    "image",
    "audio",
    "text",
    "video"
  ],
  "benchmark": {
    "image": {
      "accuracy": 0.98,
      "tpr": 1.0,
      "tnr": 0.965,
      "f1": 0.978,
      "samples": 204
    },
    "audio": {
      "accuracy": 0.852,
      "tpr": 0.85,
      "tnr": 0.855,
      "f1": 0.866,
      "samples": 142
    },
    "text": {
      "accuracy": 0.93,
      "tpr": 0.917,
      "tnr": 0.947,
      "f1": 0.936,
      "samples": 129
    },
    "video": {
      "accuracy": 0.9,
      "tpr": 0.833,
      "tnr": 1.0,
      "samples": 30
    },
    "overall": {
      "accuracy": 0.928,
      "samples": 475
    }
  },
  "image": {
    "n_features": 8,
    "feature_names": [
      "NYUAD_ViT",
      "ai_vs_real_ViT",
      "DFDC_ViT",
      "ai_image_det_ViT",
      "fft_slope",
      "hf_ratio",
      "Bombek1_SigLIP2_DINOv2",
      "SPAI_CVPR2025"
    ],
    "scaler_mean": [
      0.46721075337286583,
      0.4332848905084707,
      0.34848470501282125,
      0.7513610315914312,
      -2.7428234702735845,
      1.4757695660114816e-05,
      0.47213903127932083,
      0.5310949190042461
    ],
    "scaler_scale": [
      0.4562829992667211,
      0.4653274721438903,
      0.2594560381028844,
      0.2566914952700282,
      0.31761878154208484,
      1.745336794888413e-05,
      0.4468171423032323,
      0.4707389622737817
    ],
    "lr_coef": [
      0.6488963010751596,
      0.19470730198227582,
      0.3669096091179738,
      -1.1058065882150858,
      -0.47635552888598026,
      -0.015401252102331365,
      2.5029078795863406,
      1.237011726618108
    ],
    "lr_intercept": -0.7403570533419102,
    "hf_models": [
      "Nahrawy/AIorNot",
      "haywoodsloan/ai-image-detector-deploy",
      "prithivMLmods/Deep-Fake-Detector-Model",
      "Organika/sdxl-detector",
      "Bombek/BombDetector-v3-SigLIP2-DINOv2-LoRA"
    ],
    "notes": "Bombek1 is #1 feature (coef=+2.50). SPAI uses local repo."
  },
  "audio": {
    "n_features": 9,
    "feature_names": [
      "DavidCombei_wav2vec2",
      "Gustking_wav2vec2",
      "mo_thecreator_deepfake",
      "spectral_flatness",
      "centroid_mean",
      "centroid_std",
      "zero_crossing_rate",
      "spectral_rolloff",
      "DF_Arena_1B"
    ],
    "scaler_mean": [
      0.5667607612050348,
      0.2773010993612484,
      0.23310774392822925,
      0.03141037016224877,
      1807.2398348786571,
      897.18004887457,
      0.12301036345108962,
      6620.40736210088,
      0.5433762406366287
    ],
    "scaler_scale": [
      0.48680867334512096,
      0.29197482864644153,
      0.4211570130989059,
      0.024618810573647662,
      459.40344999868597,
      394.8528855416117,
      0.046570088698838365,
      829.6553459300637,
      0.4155082795685684
    ],
    "lr_coef": [
      0.7845433297452213,
      -0.25601227158569434,
      0.38715143588917217,
      0.5305971113288093,
      0.14191280089652655,
      1.7648106776858394,
      -1.6174243839603224,
      -1.09787021389514,
      1.092684667819162
    ],
    "lr_intercept": 0.39250921446958165,
    "hf_models": [
      "DavidCombei/wav2vec2-xls-r-1b-deepfake",
      "Gustking/ai_voice_detect_xlsr",
      "mo-thecreator/Deepfake-audio-detection",
      "Aloukik21/DF_Arena_1B"
    ]
  },
  "text": {
    "n_features": 8,
    "feature_names": [
      "binoculars_score",
      "roberta_chatgpt_score",
      "perplexity_norm",
      "avg_sentence_length",
      "lexical_diversity",
      "hapax_ratio",
      "avg_word_length",
      "fakespot_roberta_score"
    ],
    "scaler_mean": [
      1.1353826005329457,
      0.33250804246780497,
      -0.48164806951384675,
      5.916446148470062,
      0.6490103211442594,
      0.5124573713819743,
      5.220866125485708,
      0.6364287314816944
    ],
    "scaler_scale": [
      0.19535976595611237,
      0.45007809250809544,
      0.21119484430166974,
      1.1937958293169302,
      0.19352867829552858,
      0.21389850106439456,
      1.2135677101079925,
      0.43094435530407293
    ],
    "lr_coef": [
      -0.6243579398646565,
      0.389259232075374,
      -0.5040499517552531,
      -0.21291399657541557,
      -0.08360375807827485,
      -0.014109874794709326,
      0.22446151217916235,
      1.2266905154327146
    ],
    "lr_intercept": 0.1964292008569683,
    "hf_models": [
      "tiiuae/falcon-7b",
      "tiiuae/falcon-7b-instruct",
      "roberta-base-openai-detector",
      "fakespot-ai/roberta-base-fakespot-detector"
    ],
    "notes": "Binoculars uses falcon-7b pair. fakespot is #1 feature (coef=+1.23)."
  },
  "video": {
    "method": "frame_sampling + audio_extraction",
    "frame_detector": "image meta-classifier on sampled frames",
    "audio_detector": "audio meta-classifier on extracted audio (ffmpeg)",
    "combining": "50/50 equal weight of video_prob and audio_prob",
    "default_num_frames": 8,
    "notes": "Video = image detection on frames + audio detection on extracted track"
  }
}