Blame for implementation-notes.md

# Forensics-XLSR-Wild — implementation notes

Plan: `speech-spoof-bench/docs/plans/2026-11-01-nine-model-batch-arena-submission.md` (model #1).

## Source
- HF `eliya/forensics_0.3B_xlsr_wild_deepfake_classifier` @ `0cbc87b3fbd683ff8c8c961bcf244a8d5f1bcf02`
  (the old id `eliya/forensics_0.3B_xlsr_wild` redirects). CC-BY-NC-4.0 (model-card metadata; no LICENSE
  file → `LICENSE.source` records it). Downloaded only `checkpoint_epoch_4.safetensors` (1.26 GB,
  sha256 `39e37fa5a958c2b46ebb6a5937874c36c39003906dcdd5d0aba6154bc6b2dc21`), `model.py`, `inference.py`,
  `config.json`, `README.md` into `benchmarks/_ckpt_dl/eliya__forensics_0.3B_xlsr_wild_deepfake_classifier/`
  (the 3.79 GB pickled `.pt` was not fetched). `checkpoint_epoch_4.safetensors` here is a hard link to it.
- `model.py` → `_net.py` (edits in its header). Backbone config vendored as `xlsr300m_config.json`
  (`facebook/wav2vec2-xls-r-300m` @ `1a640f32ac3e39899438a2931f9924c02f080a54`); no Hub weights loaded.
- Weight-upload commits are titled "Upload itw_winner_xlsr" (2026-09-01) → checkpoint apparently picked on
  In-the-Wild; card revisions later replaced the explicit public-dataset list with "proprietary and public".

## Wrapper decisions (`forensics_xlsr_wild.py:ForensicsXLSRWild`)
- Input = `inference.py:load_audio` after mono/16 kHz: peak-normalise whole clip `wav/(max|wav|+1e-8)`,
  then centre crop `s=(cur-n)//2` if longer than 80000, repeat-pad (`repeat(ceil)[:n]`) if shorter.
  Empty clip → zeros (not in source). No resampling. FP32, cuDNN TF32 off.
- Score = the single logit (label 1 = real, sigmoid = P(real)); higher = bona fide.
- Load: checkpoint has 589 tensors; the 4 `projection.{0,2}.{weight,bias}` (160×160, training-only head,
  not in model.py — the source loads strict=False) are dropped via an asserted allow-list, then
  `strict=True` (0 missing / 0 unexpected). The source's strict=False load also reports exactly those 4
  as unexpected and no missing keys, so nothing in the released model is lost.
- Params: 315,926,375 (316.0 M; + 518 BN buffer elements).
- **Source parity** (`source_parity.py`): authors' `model.py` (from_pretrained base + strict=False) +
  `inference.py:load_audio` (torchaudio) vs our wrapper on 12 In-the-Wild clips: max |Δ| = 0.0 (bit-identical).

## Batch sweep (`sweep.py`, 80000-sample input, 2026-11-01; logs `logs_sweep_g{2,3}_tf{0,1}.txt`)
| card | best | notes |
|---|---|---|
| RTX 4070 Ti SUPER (idx 2) | 97.5 utt/s @24 | plateau 8–64; TF32 on/off identical |
| RTX 5060 Ti (idx 3) | 32.7 utt/s @1 | ~26 at bs ≥ 4 → batch_size 1 on 5060 |

## Sanity (PyTorch, 4070 Ti SUPER)
- Full J-SPAW_LA: EER **0.188%** (2397 trials, 0 skipped), ~82 utt/s end-to-end. Very low — the
  undisclosed training data may include JSUT/JVS-like Japanese speech; not flipped.
- Full InTheWild (`results_pt/InTheWild`): EER **1.092%** (31,779 trials, 0 skipped) vs self-reported
  1.03% — reproduces. (InTheWild is n/a* via trained_on.) No per-utterance author score files exist.

## TensorRT (`trt_forensics_xlsr_wild.py`, sidecar `trt_forensics-xlsr-wild.json`, TRT 10.16.1.11)
Adapted from DPDA-GA-TCM's script; `_ExportNet` returns (B,1) and n_classes is forced to 1.
FP16 fails the ranking tolerance (build Spearman 0.978 on 64 ITW clips — the AASIST top-k node
selection is precision-sensitive), so both cards use FP32.
| card | precision | best batch | utt/s (sweep) | build parity Spearman (64 ITW) | J-SPAW EER | J-SPAW Spearman vs PT / max Δ |
|---|---|---|---|---|---|---|
| RTX 4070 Ti SUPER (idx 2) | FP32 (FP16 0.978) | 8 | 136.3 | 0.99991 | 0.1879 (=PT) | 0.999991 / 0.0091 |
| RTX 5060 Ti (idx 3) | FP32 (FP16 0.978) | 4 | 91.9 | 0.99991 | 0.1879 (=PT) | 0.999982 / 0.0323 |

Both cards pass (|ΔEER| = 0 ≤ 0.25, Spearman ≥ 0.999). End-to-end J-SPAW runner throughput: PT 4070 ~82,
TRT 4070 ~114, TRT 5060 ~64 utt/s (decode-bound). FP16 engines deleted. J-SPAW_LA is submitted from
PyTorch scores (`results_pt/`); TRT runs in `results_trt2/`, `results_trt3/`; `cmp_scores.py` compares.
Scripts: `run_pt.sh`, `run_trt_export.sh`, `run_trt_jspaw.sh <gpu>`.

## meta.yaml
- No paper block: the schema makes `paper` optional but, if present, requires arxiv_id + url + bibtex, and
  any paper_url/arxiv_id moves a system into the paper-gated tiers. As with Spectra-0, the paperless model is
  left without `paper` → Unpublished tier (HF card URL is in `code`; the card's @misc bibtex is not a paper).
- `trained_on`: DFADD, ASVspoof2021_DF, ASVspoof2019_LA, InTheWild (plan Q3). Note: the package's
  `data/submission_meta.schema.json` (used by `speech-spoof-bench submit`'s `load_meta`) does not yet allow
  `trained_on` (additionalProperties false); the v5 submission schema does. `_submit_one.py` also does not copy
  `trained_on` (nor drop a null `paper`) into the submission YAML.