Add Codecfake-W2V2-AASIST-CoTrain submission for CFAD (EER=11.3196)

submissions/codecfake-w2v2-aasist-cotrain.yaml

@@ -0,0 +1,55 @@
+schema_version: 4
+system:
+  name: Codecfake-W2V2-AASIST-CoTrain
+  slug: codecfake-w2v2-aasist-cotrain
+  description: 'W2V2-AASIST countermeasure from the Codecfake paper: a frozen wav2vec 2.0 XLS-R 300M front-end
+    (per-utterance mean/variance normalised input; hidden states after its 5th transformer layer, i.e.
+    HF hidden_states[5]) feeding an AASIST head (linear 1024->128, residual encoder, spectro-temporal
+    graph attention). Weights: the co-trained release (ASVspoof2019 LA train + Codecfake train, with the
+    paper''s CSAM sampler; the paper''s headline countermeasure). The upstream repo carries no license,
+    so the weights are linked from the authors'' GitHub repo (pinned commit), not re-hosted. FP32, first
+    64600 samples (repeat-tiled if shorter); score = log-odds of bona fide.'
+  code: https://github.com/xieyuankun/Codecfake
+  checkpoint: https://github.com/xieyuankun/Codecfake/blob/72c32e2840eef95fbeade349425fa9c5391fa11e/pretrained_model/cotrain_w2v2aasist_CSAM/anti-spoofing_feat_model.pt
+  params_millions: 76.554634
+  paper:
+    arxiv_id: '2405.04880'
+    url: https://arxiv.org/abs/2405.04880
+    bibtex: "@article{xie2024codecfake,\n  title         = {The Codecfake Dataset and Countermeasures\
+      \ for the Universally Detection of Deepfake Audio},\n  author        = {Xie, Yuankun and Lu, Yi\
+      \ and Fu, Ruibo and Wen, Zhengqi and Wang, Zhiyong and Tao, Jianhua and Qi, Xin and Wang, Xiaopeng\
+      \ and Liu, Yukun and Cheng, Haonan and Ye, Long and Sun, Yi},\n  journal       = {arXiv preprint\
+      \ arXiv:2405.04880},\n  year          = {2024},\n  eprint        = {2405.04880},\n  archivePrefix\
+      \ = {arXiv}\n}\n"
+dataset:
+  id: SpeechAntiSpoofingBenchmarks/CFAD
+  revision: 53d7855c1c378524f7b7b1030bcb6b2caa327fe6
+  split: test
+scores:
+  eer_percent: 11.31958664698319
+  n_trials: 62999
+  n_skipped: 0
+artifact:
+  scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/Codecfake-W2V2-AASIST-CoTrain/resolve/a1ccfa8268c8c0ce62ce327866be49c4c37d4e9a/.eval_results/SpeechAntiSpoofingBenchmarks/CFAD/scores.txt
+  scores_sha256: ebee88569e11f330e4a6bbd4d08efe7b06eb683b6993745874271cb7a0551c8e
+  bench_version: speech-spoof-bench==0.4.1
+reproduction:
+  reproduced_by: SpeechAntiSpoofingBenchmarks
+  reproduced_at: '2026-11-02'
+  reproduced_bench_version: speech-spoof-bench==0.4.1
+  match: inference
+submitter:
+  hf_username: korallll
+  contact: k.n.borodin@mtuci.ru
+submitted_at: '2026-11-02'
+notes: 'W2VAASIST from xieyuankun/Codecfake @ 72c32e2 (model.py; generate_score.py pipeline), checkpoint
+  pretrained_model/cotrain_w2v2aasist_CSAM/anti-spoofing_feat_model.pt (a pickled whole module; converted
+  once to a state_dict with an allow-list unpickler, strict load). Frozen XLS-R = facebook/wav2vec2-xls-r-300m
+  @ 1a640f3, truncated to its first 5 transformer layers (bit-identical to the full model''s hidden_states[5]).
+  No upstream license: weights are fetched from the pinned GitHub URL at load time, not re-hosted. Deterministic
+  first 64600 samples, shorter clips repeat-tiled (source pad_dataset). Class 0 = bona fide; score = logits[:,0]
+  - logits[:,1] (monotone in the source''s softmax[:,0]). Parity vs the authors'' shipped per-utterance
+  score files (result/ @ 72c32e2), full sets: ASVspoof2019_LA eval (71,237): EER 0.312% vs authors 0.312%,
+  per-clip Pearson (logit scale) 0.99994, Spearman 0.999995; In-the-Wild (31,779): EER 4.56% vs authors
+  4.68%, Pearson 0.9986, Spearman 0.9995 (the residual is the arena copy''s 16-bit FLAC quantisation of
+  the float source wavs: on 512 original float wavs the median |d logit| drops from 0.043 to 0.0035).'