Beyond Rank: Calibration-Aware Dependency Control for Post-Training LoRA Compression
Year: 2026
Type: software
Abstract: These materials accompany Beyond Rank: Calibration-Aware Dependency Control for Post-Training LoRA Compression by Kang Wang. Compressing a trained LoRA adapter involves more than choosing how much rank to retain. When the remaining coefficients are adjusted to compensate for removed directions, limited calibration data can make detailed dependency estimates unreliable. The accompanying study investigates this problem through calibration-aware dependency control, including hierarchical Fisher shrinkage (H1) and comparisons with simpler shrinkage alternatives. This archive provides the mathematical implementations of H1, S1, G1 and the constrained solver; experimental configurations and protocols; stored numerical results; figure source data; and machine-readable tables. Favorable, adverse and inconclusive findings are included. The package also provides a standard-library integrity verifier. After extracting the ZIP, run python -I -B code/verify_release_static.py using Python 3.10 or newer. This checks file integrity and stored-data structures without retraining models or repeating the experiments. Full experimental reproduction requires additional model, dataset, adapter, Jacobian and runtime assets, as described in REPRODUCE.md. The project-authored Python code is licensed under MIT. Project-authored results, data and documentation are licensed under CC BY 4.0. File-specific licensing and third-party exceptions are detailed in LICENSE, manifests/license_scope.json and THIRD_PARTY_NOTICES.md. The formal generative-AI use declaration is provided in AI_USE.md. Source repository: https://github.com/Kane11378/calibration-aware-lora-compression-repro GitHub release: https://github.com/Kane11378/calibration-aware-lora-compression-repro/releases/tag/v1.0.1
Author: Wang, Kang
Institution: Taiyuan University of Technology
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7221321663
