Add DPDA-GA-Mamba submission for LRLspoof (1-SRR=23.5060)

submissions/dpda-ga-mamba.yaml

@@ -0,0 +1,56 @@
+schema_version: 4
+system:
+  name: DPDA-GA-Mamba
+  slug: dpda-ga-mamba
+  description: 'DPDA-GA-Mamba: the DPDA-GA training recipe (dual-path data-augmented training with PCGrad
+    gradient alignment between the clean and RawBoost-augmented paths) applied to the XLSR-Mamba architecture
+    -- a fully fine-tuned wav2vec 2.0 XLS-R 300M front-end (last layer) projected to 144-d (Linear + BatchNorm
+    + SELU), followed by a dual-column bidirectional Mamba (v1 selective SSM) back-end: 6 forward + 6
+    backward Mamba blocks with fused add+RMSNorm, per-direction attention pooling, concatenation, linear
+    fusion and a 2-way head. Official pcgrad/xlsr_mamba checkpoints (trained on ASVspoof2019 LA train),
+    5-checkpoint weight average as computed by the authors'' main.py. Not to be confused with XLSR-MamBo
+    (arXiv 2601.02944).'
+  code: https://github.com/ductuantruong/dpda_ga
+  checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/DPDA-GA-Mamba/blob/b083ed150a0eb6c5a3dba74f38795a4531baf8bb/dpda_ga_mamba_avg5.pth
+  params_millions: 319.3275
+  paper:
+    arxiv_id: '2509.20682'
+    url: https://arxiv.org/abs/2509.20682
+    bibtex: "@inproceedings{dpda_ga,\n  author    = {Truong, Duc-Tuan and Liu, Tianchi and Li, Junjie\
+      \ and Tao, Ruijie and Lee, Kong Aik and Chng, Eng Siong},\n  title     = {Addressing Gradient Misalignment\
+      \ in Data-Augmented Training for Robust Speech Deepfake Detection},\n  booktitle = {ICASSP 2026\
+      \ - 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},\n \
+      \ year      = {2026},\n  pages     = {16252--16256},\n  doi       = {10.1109/ICASSP55912.2026.11464739}\n\
+      }\n"
+dataset:
+  id: lab260/LRLspoof
+  revision: ac76f0e5472fdb68a6726e1fa94a148316e76916
+  split: test
+scores:
+  srr_complement: 23.506002672966464
+  n_trials: 1304455
+  n_skipped: 286
+calibration:
+  source_dataset: SpeechAntiSpoofingBenchmarks/DeepVoice
+  threshold: -3.505877140127387
+artifact:
+  scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/DPDA-GA-Mamba/resolve/676969850b349b6753a09b2036035ed2c44c21e0/.eval_results/lab260/LRLspoof/scores.txt
+  scores_sha256: 5afd3cc08eda0712df3479cfab24aa30966094db03fed9d53066ac97cb1e32c5
+  bench_version: speech-spoof-bench==0.4.1
+reproduction: {}
+submitter:
+  hf_username: korallll
+  contact: k.n.borodin@mtuci.ru
+submitted_at: '2026-11-01'
+notes: 'DPDA-GA (Truong et al., ICASSP 2026; arXiv 2509.20682), MIT, github.com/ductuantruong/dpda_ga;
+  checkpoints HF tuantd/dpda_ga @ 531554e0a1c28d6a6ccf22a355281e3d3acefeb5 pcgrad/xlsr_mamba/best/best_{0..4}.pth.
+  Weight average reproduces main.py --average_model True --n_average_model 5 EXACTLY, including its state_dict()
+  aliasing bug: model.state_dict() shares storage with the parameters, so loading best_1 overwrites the
+  accumulator before the first sum and the result is (2*best_1 + best_2 + best_3 + best_4)/5 (best_0 is
+  dropped) -- that is what the authors evaluated; built by make_avg5.py. Strict load. Deterministic first-66800-sample
+  window (4.175 s), shorter clips repeat-tiled (source utils.pad); no resampling or normalisation; FP32;
+  bona-fide score = logits[:, 1] (source main.py). Deviations: Mamba selective scan, causal depthwise
+  conv and RMSNorm in plain PyTorch (same math as the mamba_ssm CUDA/Triton kernels; chunked scan == selective_scan_ref
+  within 1e-4) on every GPU -- mamba_ssm/causal_conv1d are not installed and Triton is unreliable on sm_120;
+  the source''s duplicate (unused) XLS-R forward and its extract_feat train() call are removed. Not to
+  be confused with XLSR-MamBo (arXiv 2601.02944, a different model).'