Xiaolong-Han/w2t-celeba-classifier

W2T CelebA-LoRA Classifier

This repository contains the CelebA-LoRA attribute classifier checkpoint used for the W2T paper:

• Paper: W2T: LoRA Weights Already Know What They Can Do
• Repo: Weight2Token

Correct Source

This release corresponds to the Surrey experiment:

• Checkpoint: /mnt/fast/nobackup/users/xh00542/GL/best_model_celeba.pth
• Test log: /mnt/fast/nobackup/users/xh00542/GL/celeba_new/job_2057633_10.txt
• Checkpoint SHA256: 603f91a1b911842164c8ab85e7d81ee7c61c0cdf2115b3714380ed1bf57d6ad4

Test Metrics

Metrics are from job_2057633_10.txt.

Metric	Value
Loss	0.37117
Accuracy	0.90637
Macro-F1	0.50377
Micro-F1	0.75015
Mean AUROC	0.89642
Mean AUPRC	0.59949

The paper reports the CelebA-LoRA W2T row in percent as Macro-F1 50.38, Micro-F1 75.02, and AUROC 89.64.

Files

• best_model_celeba.pth: classifier checkpoint.
• job_2057633_10.txt: original Surrey training/test log.
• test_metrics.json: parsed test metrics.
• manifest.json: source paths, hashes, and release metadata.

Citation

@article{han2026w2t,
  title   = {W2T: LoRA Weights Already Know What They Can Do},
  author  = {Han, Xiaolong and Neri, Ferrante and Jiang, Zijian and Wu, Fang and Ye, Yanfang and Yin, Lu and Wang, Zehong},
  journal = {arXiv preprint arXiv:2603.15990},
  year    = {2026}
}