Qwen3.5 Audio Deepfake Detection Models

These models are checkpoints from experiments on audio deepfake detection using Qwen3.5.

Model Naming

Model names follow this general structure:

Qwen3.5_<size>_<adaptation>_<seed>_<training_setup>

Components

• 0.8B / 4B — Qwen3.5 model size.
• lora — the Qwen3.5 backbone is adapted using LoRA.
• no_lora — the Qwen3.5 backbone is frozen.
• 2 / 42 / 240 — random seed.
• binaryHIRSDD — binary training on HIR-SDD.
• binaryLA19 — binary training on LA19.
• binaryHIRSDD_and_LA19 — binary training on HIR-SDD and LA19.
• reasoningHIRSDD — reasoning-based training on HIR-SDD.
• explanationHIRSDD — explanation-based training on HIR-SDD.
• stage2 / stage3_GRPO — subsequent training stages, including GRPO.

Examples

Qwen3.5_0.8B_no_lora_2_reasoningHIRSDD

→ Qwen3.5 0.8B, frozen backbone, seed 2, reasoning training on HIR-SDD.

Qwen3.5_0.8B_lora_42_binaryHIRSDD_and_LA19

→ Qwen3.5 0.8B, LoRA adaptation, seed 42, binary training on HIR-SDD + LA19.

Qwen3.5_0.8B_lora_42_reasoningHIRSDD_stage2_HIRSDD_GRPO

→ Qwen3.5 0.8B, LoRA adaptation, seed 42, followed by an additional GRPO training stage.

The models use an audio encoder and projector together with the Qwen3.5 backbone; the naming convention above describes the main language-model and training configuration.