model.json

{
  "name": "FAD-MoE",
  "package_dir": ".",
  "entry_point": ["models.moe_research.w2v2_moe_fz24_aasist.Model", "utils.tools.tools.pad"],
  "checkpoint": {
    "backbone": {
      "repo": "facebook/wav2vec2-xls-r-300m",
      "filename": "pytorch_model.bin",
      "size_mb": 1270,
      "download": "Wav2Vec2Model.from_pretrained('facebook/wav2vec2-xls-r-300m') at startup -- replaces original repo's dead hardcoded local path '/data2/wzydata/wav2vec2-xls-r-300m'; same backbone, confirmed against pytorch_model.bin on HF Hub"
    },
    "moe_aasist_head": {
      "location": "committed directly to Space repo at checkpoints/fz24_moe_aasist.ckpt",
      "source": "FAD-MoE-repo icassp branch, a_l/noft_24/2_4_128/version_1/checkpoints/best_model-epoch=11-dev_eer=0.3560-loss=0.0015.ckpt",
      "size_mb": 99,
      "format": "pytorch-lightning checkpoint (state_dict key, load with weights_only=False), trained with moe_topk=2, moe_experts=4 (96 experts total, 4 per wav2vec2 layer), moe_exp_hid=128 -- matches paper defaults"
    }
  },
  "note": "Source repo's default branch (main) is an unrelated toy AASIST baseline -- the MoE fusion work this paper describes lives on the icassp branch (README says so explicitly). model.json/app.py are built against icassp."
}