Add Fake-Mamba submission for LRLspoof (1-SRR=15.0815)

submissions/fake-mamba.yaml

@@ -0,0 +1,61 @@
+schema_version: 4
+system:
+  name: Fake-Mamba
+  slug: fake-mamba
+  description: 'Fake-Mamba (PN-BiMamba): a fully fine-tuned wav2vec 2.0 XLS-R 300M front-end (last transformer
+    layer) projected to 144-d (BatchNorm + SELU), followed by pre-norm bidirectional Mamba blocks (one
+    shared Mamba scans the sequence forward and time-reversed, outputs summed, LayerNorm + GELU feed-forward,
+    residual), attention pooling over time and a linear 2-way head. Trained on ASVspoof2019 LA train (RawBoost).
+    Weights: the authors'' release Exp65 (12 PN-BiMamba blocks, not the paper''s 7-block configuration),
+    5-best checkpoint average exactly as their eval script computes it, (2*best_1 + best_2 + best_3 +
+    best_4)/5 (a state_dict aliasing bug drops best_0); reproduces the authors'' In-the-Wild scores (r
+    = 0.99996, EER 5.83% vs 5.85% reported). The upstream code and weights carry no license, so the weights
+    are linked from the authors'' Hugging Face repo, not re-hosted. FP32, deterministic first-66800-sample
+    window (repeat-padded).'
+  code: https://github.com/xuanxixi/Fake-Mamba
+  checkpoint: https://huggingface.co/xxuan-speech/fake-mamba-models/tree/cd99e0efb4d4f3f6dfe610ae8e27de9a73d6bf41/models/Exp65/best
+  params_millions: 321.2901
+  paper:
+    arxiv_id: '2508.09294'
+    url: https://arxiv.org/abs/2508.09294
+    bibtex: "@inproceedings{xuan2025fakemamba,\n  title     = {Fake-Mamba: Real-Time Speech Deepfake Detection\
+      \ Using Bidirectional Mamba as Self-Attention's Alternative},\n  author    = {Xuan, Xi and Zhu,\
+      \ Zimo and Zhang, Wenxin and Lin, Yi-Cheng and Kinnunen, Tomi},\n  booktitle = {Proc. IEEE Automatic\
+      \ Speech Recognition and Understanding Workshop (ASRU)},\n  year      = {2025},\n  eprint    = {2508.09294},\n\
+      \  archivePrefix = {arXiv}\n}\n"
+dataset:
+  id: lab260/LRLspoof
+  revision: ac76f0e5472fdb68a6726e1fa94a148316e76916
+  split: test
+scores:
+  srr_complement: 15.081481004379038
+  n_trials: 1304455
+  n_skipped: 286
+calibration:
+  source_dataset: SpeechAntiSpoofingBenchmarks/DeepVoice
+  threshold: 1.7670095430508475
+artifact:
+  scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/Fake-Mamba/resolve/1a126e74ef942458e3d787072811422135060085/.eval_results/lab260/LRLspoof/scores.txt
+  scores_sha256: 9ab2601702790a1b19e64bdd556617aa7433dd875539eb46331ac1bbb19a90ac
+  bench_version: speech-spoof-bench==0.4.1
+reproduction:
+  reproduced_by: SpeechAntiSpoofingBenchmarks
+  reproduced_at: '2026-11-01'
+  reproduced_bench_version: speech-spoof-bench==0.4.1
+  match: inference
+submitter:
+  hf_username: korallll
+  contact: k.n.borodin@mtuci.ru
+submitted_at: '2026-11-01'
+notes: 'Fake-Mamba from xuanxixi/Fake-Mamba @ e966d2f (model.py:Fake_Mamba + conformer00.py:PN_BiMambas),
+  checkpoint xxuan-speech/fake-mamba-models @ cd99e0e models/Exp65/best/best_{1..4}.pth averaged as SITW-DF_test.py
+  (--average_model True --n_average_model 5) actually computes it: sd = model.state_dict() aliases the
+  params, so the result is (2*b1+b2+b3+b4)/5 (maintainer decision 2026-10-30: use the weights the authors
+  evaluated). Averaged at load time; strict load, depth 12. No upstream license: weights not re-hosted.
+  Deterministic first-66800-sample window, shorter clips repeat-tiled (source utils.pad). score = logits[:,
+  1] (class 1 = bona fide; source eval scripts write out[:, 1]). Deviations: Mamba selective scan in plain
+  PyTorch (chunked parallel prefix scan; equals mamba_ssm''s sequential reference to ~1e-6) instead of
+  the CUDA kernels; forward/backward directions batched through the shared Mamba in one call; the source''s
+  extract_feat self.model.train() on device mismatch removed. Parity on the full In-the-Wild set vs the
+  authors'' score file (EER 5.85%): these weights give EER 5.83%, per-clip Pearson r = 0.99996 (single
+  best.pth: 5.97%, r = 0.989; true mean of best_0..4: 5.82%, r = 0.9992).'