frednamfred/adqa_nam_qwen3_omni_lora

ADQA Nam Qwen3-Omni LoRA Adapters

This repository contains the LoRA adapters used for the DCASE 2026 Task 5 Audio-Dependent Question Answering submission, Learning from Audio-Dependency Errors: Data Curation Strategies Based on Model Confusion Patterns in Audio Question Answering.

Code and reproduction scripts are available at https://github.com/frednam93/adqa_nam. This repository does not include DCASE audio data, generated evaluation outputs, or the base model weights. Obtain the official DCASE Task 5 data and the base model separately.

Base Model

• Qwen/Qwen3-Omni-30B-A3B-Instruct

Included Adapters

Subfolder	Description
system1_train_only_empty5_2k	Train-only strong audio-dependent data plus 5% empty-audio unknown negatives, checkpoint 2000.
system2_train_dev_empty5_2k	Train+dev strong audio-dependent data plus 5% empty-audio unknown negatives, checkpoint 2000.
system3_train_dev_empty2p5_3k	Train+dev strong audio-dependent data plus 2.5% empty-audio unknown negatives, checkpoint 3000.

The submitted ensemble combines the three systems at prediction level. Parse repair / response normalization was performed separately in the code repository and is not included as model weights here.

Loading Example

from transformers import AutoProcessor
from peft import PeftModel

base_model = "Qwen/Qwen3-Omni-30B-A3B-Instruct"
adapter_repo = "frednamfred/adqa_nam_qwen3_omni_lora"
subfolder = "system1_train_only_empty5_2k"

processor = AutoProcessor.from_pretrained(base_model, trust_remote_code=True)
# Load the Qwen3-Omni model with the model class used in your local Qwen3-Omni setup,
# then attach the LoRA adapter:
model = PeftModel.from_pretrained(model, adapter_repo, subfolder=subfolder)

For the exact inference, judging, and submission packaging flow, see the GitHub repository.

License and Terms

The adapter weights inherit restrictions from the base Qwen3-Omni model and the DCASE challenge data terms. The DCASE dataset is not redistributed here.