URARI: Utility Rank Allocation and Residual Initialization — Implementation

Abstract

This record provides the implementation of URARI (Utility Rank Allocation and Residual Initialization), a parameter-efficient fine-tuning method for code large language models. URARI combines task and semantic gradient signals to guide rank allocation under a fixed parameter budget and initializes additional low-rank dimensions using residual information outside the existing low-rank subspace. The package implements the URARI workflow, including calibration sample construction, gradient collection, utility estimation, rank allocation, residual initialization, and low-rank fine-tuning. It also provides an EvalPlus-based evaluation pipeline for HumanEval, HumanEval+, MBPP, and MBPP+. The archive contains source code, configuration files, dependency requirements, training and evaluation entry points, documentation, and lightweight tests. Model and dataset sources, reference ranks, random seeds, and training parameters are configurable. Installation, configuration, training, and evaluation instructions are provided in the README. The associated study uses DeepSeek-Coder-1.3B-Base and Qwen2.5-Coder-3B-Base. Experimental evaluation results are available separately at https://doi.org/10.5281/zenodo.23314754.