CueSpace Checkpoints

Published test weights for CueSpace — Question-Guided Structured Cue Modeling and Adaptive Fusion for Audio-Visual Question Answering.

Files

File	Dataset	--dataset	Reported test accuracy
mavqa.pt	MUSIC-AVQA	mavqa	79.22% (7232/9129)
mavqa_r.pt	MUSIC-AVQA-R	mavqa_r	same weights as mavqa.pt
mavqa_v2_balance.pt	MAVQA-v2 balance	mavqa_v2 --v2-split balance	78.49%
mavqa_v2_bias.pt	MAVQA-v2 bias	mavqa_v2 --v2-split bias	78.59%
valor32k_mcq.pt	Valor32k-AVQA MCQ	valor32k	63.03%
avqa_mcq.pt	AVQA MCQ	avqa --mcq	91.33% (15348/16805)

Download

pip install -U huggingface_hub
hf download chelili/CueSpace --local-dir ./checkpoints

Usage (CueSpace repo)

git clone https://github.com/chelilia/Cuespace.git
cd CueSpace
pip install -r requirements.txt
# prepare data/ + ckpt/ (CLIP/AST) locally — see README

python test.py --dataset mavqa --weight ./checkpoints/mavqa.pt --gpu 0
python test.py --dataset valor32k --weight ./checkpoints/valor32k_mcq.pt --gpu 0
python test.py --dataset avqa --mcq --weight ./checkpoints/avqa_mcq.pt --gpu 0

Citation

@article{cuespace2026,
  title={CueSpace: Question-Guided Structured Cue Modeling and Adaptive Fusion for Audio-Visual Question Answering},
  author={...},
  year={2026}
}