Add NII-SLT26-Qwen2Audio-Gen submission for PyAra (EER=24.2295)

submissions/nii-slt26-qwen2audio-gen.yaml

@@ -0,0 +1,52 @@
+schema_version: 4
+system:
+  name: NII-SLT26-Qwen2Audio-Gen
+  slug: nii-slt26-qwen2audio-gen
+  description: 'qwen2audio-gen from the NII SLT 2026 study "What Do Deepfake Audio Detectors Hear?" (Ge,
+    Wang, Yamagishi): Qwen2-Audio-7B-Instruct with LoRA (r=16, alpha=32) on the q/v projections of all
+    32 audio-encoder and 32 LLM attention layers, trained generatively on CompSpoof V2 (speech + environmental
+    sound mixtures; any fake component = fake) to answer the chat prompt "Listen carefully to this recording
+    ... Classify as either real or fake." with the word real or fake. Score = logit("real") - logit("fake")
+    of the first answer token (verbalizer head, higher = bona fide), as the authors'' scorer reads it.
+    LoRA merged at load; BF16 backbone split over two GPUs, the two answer logits in FP32. Input: first
+    4 s (64600 samples), no repeat-tiling.'
+  code: https://github.com/nii-yamagishilab/AntiDeepfake-AudioLLM/tree/2d445b525bb52518448b353fe1fee51d5f815af9
+  checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/NII-SLT26-Qwen2Audio-Gen/blob/93cf1f9eb081f9c4a468abff9ba1fbe4e42107f6/weights/adapter.safetensors
+  params_millions: 8397.094912
+  paper:
+    url: https://arxiv.org/abs/2610.17617
+    arxiv_id: '2610.17617'
+    bibtex: "@inproceedings{ge2026what,\n  title     = {What Do Deepfake Audio Detectors Hear? From Self-Supervised\
+      \ Models to Audio LLMs},\n  author    = {Ge, Wanying and Wang, Xin and Yamagishi, Junichi},\n  booktitle\
+      \ = {IEEE Spoken Language Technology Workshop (SLT)},\n  year      = {2026}\n}\n"
+dataset:
+  id: SpeechAntiSpoofingBenchmarks/PyAra
+  revision: 40167badd6d8a289f02054761ee5411f49227b87
+  split: test
+scores:
+  eer_percent: 24.229494130036272
+  n_trials: 201778
+  n_skipped: 0
+artifact:
+  scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/NII-SLT26-Qwen2Audio-Gen/resolve/d71ba691ca620b553436cb38af76250c996b58e6/.eval_results/SpeechAntiSpoofingBenchmarks/PyAra/scores.txt
+  scores_sha256: 482cee568369d2cd620d62e8f5d06d34f020e86e21a8a06256c84af3721fe4ea
+  bench_version: speech-spoof-bench==0.4.1
+reproduction: {}
+submitter:
+  hf_username: korallll
+  contact: k.n.borodin@mtuci.ru
+submitted_at: '2026-11-13'
+notes: 'Checkpoint: nii-yamagishilab/slt2026-checkpoints @ ae9d2bc72b670e49513f5bd125bf679095bf4b92, checkpoints.tar.gz
+  -> qwen2audio-gen.ckpt (sha256 e725b1620431078960366d8593288ccdbee3b6440c3b7ecf6282da94734bfd67; CC-BY-NC-4.0),
+  a pickled bf16 training checkpoint of the whole peft-wrapped model (loaded with weights_only=True).
+  Its 876 base tensors are bit-identical to Qwen/Qwen2-Audio-7B-Instruct @ 0a095220c30b7b31434169c3086508ef3ea5bf0a
+  (Apache-2.0, linked), so only the LoRA (+ head) is re-hosted as weights/adapter.safetensors (sha256
+  578d5b7c7d378a0b1de39e923f644c654c0ba5e004df35ecb96b245b3bd6e2de). Strict load: every base tensor loaded,
+  every adapter tensor used (asserted). Deviations: Arena 4 s cap (authors: random 30 s crop; CompSpoof
+  V2 clips are 4 s); LoRA merged into bf16 weights and the head computed in FP32 (authors: unmerged LoRA,
+  bf16 head). Parity vs the authors'' collator + head code on 17 J-SPAW_LA clips: max |d score| = 0.214
+  (bf16 level). Training data: CompSpoof V2 (AnyFakeNoOriginal protocol); its train sources include LibriTTS
+  and ASVspoof 5 (spoof) with splits unstated (card, ESDD2 overview arXiv 2606.10791) -> whole corpora
+  -> ASVspoof5, SONAR, CD-ADD, LibriSeVoc marked trained_on (n/a*); also CommonVoice, MLAAD, english-conversation-corpus
+  (no Arena dataset). No self-reported numbers outside CompSpoof V2. PyTorch only (7B, no TensorRT). J-SPAW_LA
+  scored with this wrapper (EER 20.413%).'