# DeepFense Paper Scores

Canonical, reproducible score bundle for the DeepFense camera-ready paper.

**Hugging Face:** [DeepFense/prediction_scores](https://huggingface.co/datasets/DeepFense/prediction_scores)

---

## Directory layout

```
scores/
├── README.md
├── {train_recipe}/                              # dataset the model was TRAINED on
│   ├── {backend}/                               # AASIST | MLP | Nes2Net | TCM
│   │   └── {frontend}/                          # Wav2Vec2 | HuBERT | WavLM | EAT
│   │       └── seed{2|42|240}/
│   │           └── {eval_benchmark}/            # held-out TEST set
│   │               ├── predictions.txt          # per-utterance scores (TSV)
│   │               └── metrics.json             # EER, ACC, F1, …
│   └── _summaries/{eval_benchmark}.json         # optional cross-architecture tables
└── bias_fairness/
    └── {accent|emotions|gender|language|quality}/
        └── {eval_benchmark}/
            └── {train_recipe}/{backend}/{frontend}/seed{N}/
                └── utterances.txt               # scores + subgroup metadata (TSV)
```

### Example

From checkpoint `DeepFense_ADD23_Wav2Vec2_TCM_NoAug_Seed240` evaluated on `add22_test_track1`:

```
ADD23/TCM/Wav2Vec2/seed240/add22_test_track1/predictions.txt
ADD23/TCM/Wav2Vec2/seed240/add22_test_track1/metrics.json
```

---

## Naming conventions

| Token | Canonical form | Notes |
|-------|----------------|-------|
| **Train recipe** | `ASV5`, `ASV19`, `ADD23`, `CodecFake`, `HABLA`, `PartialSpoof` | Training dataset (not the eval set). `ASV5` = trained on ASVspoof 5; do not confuse with eval `asvspoof5_test`. |
| **Frontend** | `Wav2Vec2`, `HuBERT`, `WavLM`, `EAT` | Always PascalCase; `Hubert` → `HuBERT`. |
| **Backend** | `AASIST`, `MLP`, `Nes2Net`, `TCM` | Uppercase acronym. |
| **Seed** | `seed2`, `seed42`, `seed240` | Three seeds per recipe. |
| **Eval benchmark** | lowercase snake_case | e.g. `asvspoof5_test`, `asvspoof2019_la_eval`, `mlaad_final`, `add22_test_track1`. |

### Eval benchmarks (20)

`add22_test_track1`, `add22_test_track3`, `add23_test_R1`, `add23_test_R2`, `asvspoof2019_la_eval`, `asvspoof21_df_eval`, `asvspoof21_la_eval`, `asvspoof5_test`, `codecfake_eval`, `ctrsvdd_eval`, `fakemusiccaps_eval`, `habla_test`, `itw_eval`, `mlaad_final`, `odss_test`, `partialedit_eval`, `partialspoof_eval`, `replaydf_all_eval`, `spoofceleb_eval`

---

## File formats (`.txt` / `.json` only)

### `predictions.txt` — clip-level (TSV)

```
utterance_id	label	score_spoof	score_bonafide
LA_E_12345	0	-2.14895	3.14895
LA_T_67890	1	4.37140	-3.37140
```

- `label`: `0` = spoof, `1` = bonafide  
- LLR = `score_bonafide − score_spoof`

### `metrics.json`

Per-run aggregated metrics (EER, ACC, F1, confidence intervals).

### `bias_fairness/.../utterances.txt`

Per-utterance scores with subgroup columns (gender, accent, NISQA quality, etc.).  
Score columns: `score_spoof`, `score_bonafide` (renamed from legacy `class0`/`class1`).