gonnerthetooner/deepfake-temporal-moe

Temporal Deepfake MoE

This repository stores checkpoints and artifacts produced by temporal_deepfake_moe_hf_colab.py.

System

• Micro expert: short-range frame-difference artifact detector.
• Mid expert: CLIP frame encoder + temporal Transformer.
• Long expert: sparse CLIP frame encoder + temporal Transformer.
• Extra-long expert: optional inference-escalation expert.
• Spatial detector: external detector wrapper or stub.
• Lip-sync model: external audio-video wrapper or stub.
• Fusion: Transformer over variable expert tokens with reliability weighting.

Runtime

• Device: cuda
• GPU: Tesla T4
• VRAM GB: 14.6
• Precision: fp16
• Embedding dim: 256

Metrics

{
  "stage": "stage3",
  "temperature": 0.9831728339195251,
  "validation": {
    "n": 9,
    "loss": 0.6016786694526672,
    "accuracy": 0.6666666666666666,
    "precision": 0.0,
    "recall": 0.0,
    "f1": 0.0,
    "roc_auc": 0.9444444444444444,
    "confusion_matrix": [
      [
        6,
        0
      ],
      [
        3,
        0
      ]
    ],
    "per_dataset": {
      "unknown": {
        "n": 9,
        "loss": 0.0,
        "accuracy": 0.6666666666666666,
        "precision": 0.0,
        "recall": 0.0,
        "f1": 0.0,
        "roc_auc": 0.9444444444444444,
        "confusion_matrix": [
          [
            6,
            0
          ],
          [
            3,
            0
          ]
        ]
      }
    }
  },
  "time": "2026-06-09T17:00:25Z"
}

Notes

The training pipeline is streaming-based. It builds JSONL metadata manifests and decodes sampled frames one video at a time. It does not depend on Google Drive.