Add HoliAntiSpoof submission for CFAD (EER=22.7833)

submissions/holiantispoof.yaml

@@ -0,0 +1,56 @@
+schema_version: 4
+system:
+  name: HoliAntiSpoof
+  slug: holiantispoof
+  description: 'HoliAntiSpoof (Xu et al., 2026): an audio LLM for holistic speech anti-spoofing. The Qwen2.5-Omni-7B
+    thinker (Whisper-style audio encoder + Qwen2 7B LLM) is fine-tuned to emit a structured JSON analysis
+    of a clip (real/fake, spoofing method, spoofed regions, semantic influence): DoRA r=64 on the LLM
+    self-attention q/k/v/o projections plus a fully fine-tuned audio encoder. Arena wrapper: one forward
+    pass over the authors'' prompt with the fixed answer prefix {"real_or_fake": " and score = logit("real")
+    - logit("fake") of the next token (higher = bona fide; monotone in the authors'' p(real)). DoRA merged
+    into the base weights at load; BF16, layers split over two GPUs; the two answer logits are computed
+    in FP32. Input: first 4 s (64600 samples) of the clip, no repeat-tiling, clips under 1 s zero-padded
+    to 1 s (as the authors do).'
+  code: https://github.com/OpenTSLab/HoliAntiSpoof/tree/cb328097c211f52546f627e44e6bb2b95178232d
+  checkpoint: https://huggingface.co/wsntxxn/HoliAntiSpoof/tree/d86a035c2c626b760ef483d2ece373eb203a9ec6
+  params_millions: 8255.263744
+  paper:
+    url: https://arxiv.org/abs/2602.04535
+    arxiv_id: '2602.04535'
+    bibtex: "@article{xu2026holiantispoof,\n  title={HoliAntiSpoof: Audio LLM for Holistic Speech Anti-Spoofing},\n\
+      \  author={Xu, Xuenan and Ren, Yiming and Liu, Liwei and Wu, Wen and Li, Baoxiang and Lu, Chaochao\
+      \ and Wang, Shuai and Zhang, Chao},\n  journal={arXiv preprint arXiv:2602.04535},\n  year={2026}\n\
+      }\n"
+dataset:
+  id: SpeechAntiSpoofingBenchmarks/CFAD
+  revision: 53d7855c1c378524f7b7b1030bcb6b2caa327fe6
+  split: test
+scores:
+  eer_percent: 22.78333333333333
+  n_trials: 62999
+  n_skipped: 0
+artifact:
+  scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/HoliAntiSpoof/resolve/a336117668153709dbae886a4b251537f3ec35c0/.eval_results/SpeechAntiSpoofingBenchmarks/CFAD/scores.txt
+  scores_sha256: e92664a74fe8f015f54469944605713da9a366cc17feb6e65b08001cf526f33c
+  bench_version: speech-spoof-bench==0.4.1
+reproduction:
+  reproduced_by: SpeechAntiSpoofingBenchmarks
+  reproduced_at: '2026-11-04'
+  reproduced_bench_version: speech-spoof-bench==0.4.1
+  match: inference
+submitter:
+  hf_username: korallll
+  contact: k.n.borodin@mtuci.ru
+submitted_at: '2026-11-04'
+notes: 'Adapter = HF wsntxxn/HoliAntiSpoof @ d86a035c2c626b760ef483d2ece373eb203a9ec6 (adapter_model.safetensors:
+  DoRA r=64, alpha=128 on self_attn.{q,k,v,o}_proj + the fine-tuned audio_tower as peft modules_to_save),
+  linked not re-hosted; base = Qwen/Qwen2.5-Omni-7B @ ae9e1690543ffd5c0221dc27f79834d0294cba00 (thinker
+  tensors from shards 1-4; vision tower, talker and token2wav not loaded). DoRA merged in fp32 with peft
+  semantics; every expected tensor loaded and every adapter tensor used (asserted). Prompt, answer prefix
+  and audio preprocessing follow the authors'' inference.py / data_qwen.py (WhisperFeatureExtractor; padding
+  only to the batch''s longest clip + 0.1 s instead of 300 s -- masked frames never reach the model; logits
+  verified identical). The 4 s cap is the maintainer''s choice (the authors train on up to 60 s). Training
+  data (paper Table 3): ASVspoof2019 LA train, PartialSpoof, SINE, WaveFake, CodecFake, PartialEdit, VCTK,
+  LJSpeech, EmoFake (27,300 clips, split unstated), SpeechFake-en, DailyTalkEdit, FakeOrReal and all 31,779
+  In-the-Wild clips -> InTheWild and EmoFake_test marked trained_on (n/a*). J-SPAW_LA scored with this
+  wrapper (EER 10.52%).'