AudioLLM-Deepfake-Detection — Results Hub

Best-run checkpoints, evaluation CSVs/JSONs, and aggregated metrics for the DeepFense AudioLLM Deepfake Detection project.

Hub repo: 01Yassine/AudioLLM-Deepfake-Detection

Contents (~283 GB)

Folder	Description	Size
Suite/	Baseline Whisper/Wav2Vec + Qwen (LoRA α=16/64/128/256, NoLoRA)	~82 GB
OpenSmile/	OpenSmile before stage	~54 GB
OpenSmile-After/	OpenSmile after stage (best overall)	~62 GB
EAT/	EAT encoder experiments	~27 GB
Weighted/	Layer-weighted fusion	~59 GB
DAC-6-Lora-Qwen0.5B/	DAC tokenizer baseline	~8 MB
SpeechTokenizer-4-Lora-Qwen0.5B/	SpeechTokenizer baseline	~18 MB
Qwen-Omni-3B-lora-full/	Qwen2.5-Omni end-to-end	~422 MB
Qwen-Omni-3B-lora-opensmile/	Qwen2.5-Omni + OpenSmile	~462 MB

Each run folder includes: best_run_meta.json, per-dataset eval CSVs, metrics JSON (with EER), and checkpoints (lora_best/, checkpoint_best.pt, etc.).

Aggregated metrics (machine-readable)

File	Description
all_results_table.json	Full nested table (89 runs): Macro F1, Accuracy, EER
all_results_table.csv	Flat CSV for spreadsheets / LaTeX
all_results_table.md	Markdown tables by experiment family

Best overall model

OpenSmile-After / Lora-256 / unfrozen / Whisper / Qwen-0.5B / α=256

• Avg Macro F1: 94.42%
• Avg Accuracy: 95.11%
• Avg EER: 5.36%
• Path: OpenSmile-After/Lora-256/unfrozen/whisper/Qwen-0.5B/

---

Results Summary (local documentation)

Unified table of best runs across all experiment families.

Metrics

Metric	Description
Macro F1	Unweighted average of Fake-class and Real-class F1 (equal weight per class)
Accuracy	Fraction of correct predictions
EER	Equal Error Rate from score_fake vs ground truth (lower is better); from metrics JSON when available

Datasets

• asv19_test — ASVspoof 2019 LA eval
• itw — In-The-Wild
• la21 — ASVspoof 2021 LA eval
• mlaad_en — MLAAD English

Averages (avg_*) are computed over evaluated datasets for each run (typically 4/4).

Experiment Families

Family	Description
Suite	Baseline Whisper/Wav2Vec + Qwen LLM (LoRA α=16/64/128/256, frozen/unfrozen, NoLoRA)
OpenSmile	OpenSmile features injected before LLM (stage 1)
OpenSmile-After	OpenSmile features injected after audio encoder (stage 2)
EAT	EAT audio encoder + Qwen
Weighted	Layer-weighted fusion of Whisper/Wav2Vec representations
DAC-6-Lora-Qwen0.5B	DAC tokenizer (6 codebooks) + Qwen-0.5B LoRA
SpeechTokenizer-4-Lora-Qwen0.5B	SpeechTokenizer + Qwen-0.5B LoRA
Qwen-Omni-3B-lora-full	Qwen2.5-Omni-3B end-to-end LoRA
Qwen-Omni-3B-lora-opensmile	Qwen2.5-Omni-3B LoRA + OpenSmile

Best Overall Models

Highest average Macro F1

• OpenSmile-After / Lora-256 / unfrozen / whisper / Qwen-0.5B / α=256 — 94.42% macro F1, 95.11% accuracy, 5.36% EER
• Path: results/OpenSmile-After/Lora-256/unfrozen/whisper/Qwen-0.5B

Lowest average EER

• OpenSmile-After / Lora-256 / unfrozen / whisper / Qwen-0.5B / α=256 — 5.36% EER, 94.42% macro F1, 95.11% accuracy
• Path: results/OpenSmile-After/Lora-256/unfrozen/whisper/Qwen-0.5B

Files

File	Format	Use
all_results_table.json	Nested JSON	Machine-readable; full per-dataset breakdown
all_results_table.csv	Flat CSV	Spreadsheet / LaTeX table generation
all_results_table.md	Markdown tables	Human-readable, grouped by family

Notes

• Some runs borrow missing eval splits (documented in JSON notes / borrowed_or_approximate).
• OpenSmile before stage: LoRA α=16 and Lora-128 only for Qwen-0.5B; NoLoRA all sizes.
• OpenSmile after stage: full LoRA α sweep (0.5B) + α=16 for 3B/7B.
• EER requires score columns in eval CSV; if metrics JSON missing, EER computed from CSV.