Add DF Arena 500M submission (eer=4.3303%, scoring-verified)

submissions/df-arena-500m.yaml

@@ -0,0 +1,48 @@
+schema_version: 4
+system:
+  name: DF Arena 500M
+  slug: df-arena-500m
+  description: RAPTOR universal anti-spoofing model. A wav2vec 2.0 XLS-R 300M self-supervised front-end
+    whose per-layer hidden states are combined by learnable attention pooling (a layer-wise sigmoid gate
+    over an attention-pooled summary), then passed through a 4-block Conformer head with a class token
+    to a 2-way classifier. FP32, deterministic first-64600-sample (~4.04 s @ 16 kHz) window, tile-repeat
+    if shorter (no random crop, no resampling). score = softmax(logits)[bonafide]; higher = more bona
+    fide. Official Speech-Arena-2025/DF_Arena_500M_V_1 checkpoint.
+  code: https://huggingface.co/Speech-Arena-2025/DF_Arena_500M_V_1
+  checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/DF_Arena_500M_V_1
+  params_millions: 436.1919
+  paper:
+    arxiv_id: '2603.06164'
+    url: https://arxiv.org/abs/2603.06164
+    bibtex: "@misc{kulkarni2026compactsslbackbonesmatter,\n  title={Do Compact SSL Backbones Matter for\
+      \ Audio Deepfake Detection? A Controlled Study with RAPTOR},\n  author={Ajinkya Kulkarni and Sandipana\
+      \ Dowerah and Atharva Kulkarni and Tanel Alumäe and Mathew Magimai Doss},\n  year={2026},\n  eprint={2603.06164},\n\
+      \  archivePrefix={arXiv},\n  primaryClass={cs.SD},\n  url={https://arxiv.org/abs/2603.06164}\n}\n"
+dataset:
+  id: SpeechAntiSpoofingBenchmarks/DECRO
+  revision: e54c29df5b81bac75e5cc294d1ab0d4db72a5c71
+  split: test
+scores:
+  eer_percent: 4.330292846855497
+  n_trials: 37314
+  n_skipped: 0
+artifact:
+  scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/DF_Arena_500M_V_1/resolve/1af1f09ebe5ec6b57f48ea6c980c90f4c349ca1e/.eval_results/SpeechAntiSpoofingBenchmarks/DECRO/scores.txt
+  scores_sha256: c1259f8f7047d778595ab4516dfedd730c6e20ebbb1319d6f1d84b7b61538fb4
+  bench_version: speech-spoof-bench==0.4.1
+submitter:
+  hf_username: korallll
+  contact: kborodin.research@gmail.com
+submitted_at: '2026-06-25'
+notes: DF Arena 500M (RAPTOR family), XLS-R 300M (wav2vec 2.0) front-end + attention layer pooling + 4-block
+  Conformer back-end. Architecture is built from the facebook/wav2vec2-xls-r-300m config, then every weight
+  is overwritten by the fine-tuned checkpoint (pytorch_model.bin). Trained on traditional + singing +
+  environmental deepfake corpora (ASVspoof 2019/2024, Codecfake, LibriSeVoc, DFADD, CTRSVDD, SpoofCeleb,
+  MLAAD, EnvSDD). The vendored forward feeds a real (B,T) batch to the SSL model (source hard-coded batch-1
+  via unsqueeze(0)); numerically identical per item in eval() (BatchNorm uses running stats). Deterministic
+  first-64600-sample window. score = output softmax prob for class 1 (bona fide); higher = more bona fide.
+reproduction:
+  reproduced_by: SpeechAntiSpoofingBenchmarks
+  reproduced_at: '2026-06-25'
+  reproduced_bench_version: speech-spoof-bench==0.4.1
+  match: scoring