Add Forensics-WavLM-Base submission for LRLspoof (1-SRR=1.9187)

#24
by korallll - opened
submissions/forensics-wavlm-base.yaml ADDED
@@ -0,0 +1,52 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ schema_version: 4
2
+ system:
3
+ name: Forensics-WavLM-Base
4
+ slug: forensics-wavlm-base
5
+ description: 'Forensics 0.3B Base speech deepfake detector (Eliya Segev, 2026): a fully fine-tuned WavLM-Large
6
+ front-end whose 24 transformer hidden states are pooled by an AASIST-style spectro-temporal graph-attention
7
+ pool (layer x time graphs, GAT dims 128/64), followed by a 3-block residual MLP classifier (320-d)
8
+ and a single logit (label 1 = real); trained with label-smoothed CE + OC-Softmax + real-anchor supervised-contrastive
9
+ losses. Authors'' checkpoint_epoch_5 as released. Input as the authors'' inference.py: peak-normalised
10
+ waveform, 5 s (80000 samples) centre crop, shorter clips repeat-padded. score = the logit (sigmoid
11
+ = P(real)). Unpublished model (HF model card only, no paper).'
12
+ code: https://huggingface.co/eliya/forensics_0.3B_base_deepfake_classifier/tree/cfd11365b02167214dc3b392378809fef6ca3cfa
13
+ checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/Forensics-WavLM-Base/blob/4202d397214764a88e91d6b69364861306c53f2d/checkpoint_epoch_5.safetensors
14
+ params_millions: 317.257863
15
+ dataset:
16
+ id: lab260/LRLspoof
17
+ revision: ac76f0e5472fdb68a6726e1fa94a148316e76916
18
+ split: test
19
+ scores:
20
+ srr_complement: 1.9186930528175439
21
+ n_trials: 1304455
22
+ n_skipped: 286
23
+ calibration:
24
+ source_dataset: SpeechAntiSpoofingBenchmarks/DeepVoice
25
+ threshold: 1.7935821369426752
26
+ artifact:
27
+ scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/Forensics-WavLM-Base/resolve/331eb04a6e2a9c9b83a8788ce62a3129db07d51d/.eval_results/lab260/LRLspoof/scores.txt
28
+ scores_sha256: ffc34ffb646b2886928b06b9d7fbc3e35bd4f2c681eef31d91720934162ca052
29
+ bench_version: speech-spoof-bench==0.4.1
30
+ reproduction:
31
+ reproduced_by: SpeechAntiSpoofingBenchmarks
32
+ reproduced_at: '2026-11-09'
33
+ reproduced_bench_version: speech-spoof-bench==0.4.1
34
+ match: inference
35
+ submitter:
36
+ hf_username: korallll
37
+ contact: k.n.borodin@mtuci.ru
38
+ submitted_at: '2026-11-09'
39
+ notes: 'Forensics-WavLM-Base = HF eliya/forensics_0.3B_base_deepfake_classifier @ cfd11365b02167214dc3b392378809fef6ca3cfa,
40
+ checkpoint_epoch_5.safetensors (sha256 803d2c57...90fb), CC-BY-NC-4.0, re-hosted unchanged. Architecture
41
+ = the repo''s model.py (byte-identical across the WavLM Forensics models); the WavLM-Large backbone
42
+ is built from the microsoft/wavlm-large config and every weight is overwritten by the checkpoint (strict
43
+ load, incl. the weight-normed pos_conv; the 4 training-only projection.* tensors, unused by model.py,
44
+ are dropped via an asserted allow-list). Preprocessing = inference.py: peak normalisation wav/(max|wav|+1e-8)
45
+ over the whole clip, then a centred 80000-sample (5 s) crop or repeat-padding; no resampling; FP32.
46
+ Score = the single output logit (higher = bona fide); verified bit-identical to the authors'' model.py
47
+ + inference.py on 20 In-the-Wild clips. Training data: the current model card says 5M+ samples from
48
+ a mix of proprietary and public data; earlier card revisions listed SpeechFake, MD-CommonVoice, DFADD,
49
+ CodecFake, ASVspoof2019-LA, EnvSDD (later ones add ASVspoof2021, CommonVoice, Aishell); splits are not
50
+ stated, so InTheWild, DFADD, ASVspoof2021_DF and ASVspoof2019_LA are marked trained_on (n/a*), as for
51
+ Forensics-XLSR-Wild. Self-reported EER: In-the-Wild 1.38%, ASVspoof2019-LA 0.26%, ASVspoof2021-LA 1.56%.
52
+ No paper -> Unpublished tier.'