arXiv subjects

Daniel J Perry

Publications and source records attributed to Daniel J Perry.

3 recordsLinked to original sources

arXiv2026

Improving Training Efficiency and Reducing Maintenance Costs via Language Specific Model Merging

Fine-tuning a task-specific multilingual large language model (LLM) involves training the model on a multilingual dataset with examples in all the required languages. Updating one or more supported languages with additional data or adding support for a new language involves retraining the model, which can be computationally inefficient and creates a severe maintenance bottleneck. Recent research on merging multilingual multitask models has shown promise in terms of improved quality, but its computational and maintenance efficiency remains unstudied. In this work, we provide the first focused analysis of this merging strategy from an efficiency perspective, evaluating it across three independent tasks. We demonstrate significant efficiency gains while maintaining parity in terms of quality: this merging approach reduces the initial training time by up to 50\%. We also demonstrate that updating an individual language and re-merging as part of model maintenance reduces training costs by more than 60\%, compared to re-training the full multilingual model. We show this on both public and proprietary industry datasets confirming that the approach works well for industrial use cases in addition to academic settings already studied in previous work.

cs.CL↗

arXiv2023

Effective Proxy for Human Labeling: Ensemble Disagreement Scores in Large Language Models for Industrial NLP

Large language models (LLMs) have demonstrated significant capability to generalize across a large number of NLP tasks. For industry applications, it is imperative to assess the performance of the LLM on unlabeled production data from time to time to validate for a real-world setting. Human labeling to assess model error requires considerable expense and time delay. Here we demonstrate that ensemble disagreement scores work well as a proxy for human labeling for language models in zero-shot, few-shot, and fine-tuned settings, per our evaluation on keyphrase extraction (KPE) task. We measure fidelity of the results by comparing to true error measured from human labeled ground truth. We contrast with the alternative of using another LLM as a source of machine labels, or silver labels. Results across various languages and domains show disagreement scores provide a better estimation of model performance with mean average error (MAE) as low as 0.4% and on average 13.8% better than using silver labels.

cs.CL↗

arXiv2020

Visualization of topology optimization designs with representative subset selection

An important new trend in additive manufacturing is the use of optimization to automatically design industrial objects, such as beams, rudders or wings. Topology optimization, as it is often called, computes the best configuration of material over a 3D space, typically represented as a grid, in order to satisfy or optimize physical parameters. Designers using these automated systems often seek to understand the interaction of physical constraints with the final design and its implications for other physical characteristics. Such understanding is challenging because the space of designs is large and small changes in parameters can result in radically different designs. We propose to address these challenges using a visualization approach for exploring the space of design solutions. The core of our novel approach is to summarize the space (ensemble of solutions) by automatically selecting a set of examples and to represent the complete set of solutions as combinations of these examples. The representative examples create a meaningful parameterization of the design space that can be explored using standard visualization techniques for high-dimensional spaces. We present evaluations of our subset selection technique and that the overall approach addresses the needs of expert designers.

cs.GR↗