Sor0ush/repro-when-shared-knowledge-hurts-spectral-over-accumulation-in-model-merging-artifacts / repro-bundle /README.md
SVC reproduction bundle
Scripts and outputs for reproducing Singular Value Calibration (SVC) from When Shared Knowledge Hurts (ICML 2026, arXiv:2602.05536).
Quick start
uv sync # or: pip install -r requirements from environment.yml
python download_checkpoints.py
uv run python verify_repro.py
What this verifies
- Claim 1: SVC direction preservation (cosine sim ≈ 1)
- Claim 5: spectral overlap / singular-value gap plots (
plot_figure3.png,plot_figure4.png) - Claim 6: layer-wise SVD calibration timing
- Synthetic forward pass through merged encoder + classification head
Not included: full Table 1/2 accuracy evaluation (needs vision datasets / NLP models).
Author repo: https://github.com/lyymuwu/SVC
Xet Storage Details
- Size:
- 746 Bytes
- Xet hash:
- 039d4f2bdc92c065e54ecf1379a3a21646a862526d15012414b8f4a44844e8f1
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.