MQ-Auditor HyperLoRA Weights

This repository contains the released MQ-Auditor pretrained weights for reference-free mask quality assessment in language-referred audio-visual segmentation.

The checkpoint corresponds to:

epochs96_lr1e-4_bs4_gradacc8_lora_r32alpha64_pos0.5_ioulosswei0

Model

MQ-Auditor takes a video clip, audio, a referring expression, a frame, and a candidate segmentation mask, then predicts mask quality attributes such as mask type, IoU, and recommended action.

The released weights are intended to be used with the MQ-Auditor codebase and MQ-RAVSBench dataset. The base LLM checkpoint and external encoders are not included in this package.

Release Contents

The public weight package should include:

adapter_config.json
adapter_model.safetensors
config.json
model.txt
model_trainable_params.txt
non_lora_trainables.bin
saved_config.json
trainer_state.json
checkpoint-960/
  config.json
  finetune_weights.bin

Intermediate epoch checkpoints and TensorBoard logs are not part of the release package.

Training Data

The model was trained on MQ-RAVSBench with:

train_test_meta_files/metadata.csv
train_test_meta_files/train_audit_only_filtered.json

null masks are used during training as empty-mask examples. They are not part of the default/reported test-time evaluation protocol.

Evaluation

Evaluation is reported on the seen and unseen MQ-RAVSBench test splits:

test_s_image_filtered.json
test_u_image_filtered.json
test_s_video_filtered.json
test_u_video_filtered.json

Reported mask types focus on non-empty candidate masks: perfect, cutout, erode, dilate, merge, and full_neg.

License

The released MQ-Auditor weights are provided for non-commercial research purposes only under CC BY-NC-SA 4.0-style terms. The weights depend on the Llama-2 base model and other pretrained encoders, so users must also comply with the applicable upstream model licenses and access terms.

Citation

@article{zhou2026audit,
  title={Audit After Segmentation: Reference-Free Mask Quality Assessment for Language-Referred Audio-Visual Segmentation},
  author={Zhou, Jinxing and Zhou, Yanghao and Wang, Yaoting and Han, Zongyan and Ma, Jiaqi and Ding, Henghui and Anwer, Rao Muhammad and Cholakkal, Hisham},
  journal={arXiv preprint arXiv:2602.03892},
  year={2026}
}

Paper: https://arxiv.org/pdf/2602.03892