SARG-QLoRA: Syntax-Aware Retrieval-Guided Parameter-Efficient Fine-Tuning for Dependency Parsing with Autoregressive Large Language Models

Reference implementation and reproducibility artefacts accompanying the paper "SARG-QLoRA: Syntax-Aware Retrieval-Guided Parameter-Efficient Fine-Tuning for Dependency Parsing with Autoregressive Large Language Models." SARG-QLoRA couples Syntax-Aware Retrieval Guidance (SARG) -- a retrieval mechanism that selects supporting exemplars by structural similarity rather than semantic similarity -- with Quantized Low-Rank Adaptation (QLoRA) fine-tuning of a 3B-parameter instruction-tuned backbone, to perform generative dependency parsing on a single commodity GPU. Output reliability is validated with a four-criterion Tree Validity Checker (TVC), applied independently of gold-reference annotations. Trained and evaluated on the Universal Dependencies English Web Treebank, the model achieves 81.35% UAS / 78.28% LAS (within 1.07 points of the specialist Stanza parser while updating only 0.75% of model parameters), 96.97% structural validity under the Tree Validity Checker, and sustains zero-shot cross-lingual transfer to Spanish, German, and French without any multilingual training data. This archive includes the paper source (LaTeX), the full experiment notebook, real model prediction outputs, result tables, and a standalone, independently-runnable script (tvc_scoring.py) that recomputes every structural-reliability figure reported in the paper directly from the model's own saved output.