Enregistré dans:
Détails bibliographiques
Auteurs principaux: Bush, Sam, DeLorenzo, Matthew, Tieu, Phat, Rajendran, Jeyavijayan
Format: Preprint
Publié: 2025
Sujets:
Accès en ligne:https://arxiv.org/abs/2505.06096
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866916817701699584
author Bush, Sam
DeLorenzo, Matthew
Tieu, Phat
Rajendran, Jeyavijayan
author_facet Bush, Sam
DeLorenzo, Matthew
Tieu, Phat
Rajendran, Jeyavijayan
contents Limitations in Large Language Model (LLM) capabilities for hardware design tasks, such as generating functional Verilog codes, have motivated various fine-tuning optimizations utilizing curated hardware datasets from open-source repositories. However, these datasets remain limited in size and contain minimal checks on licensing for reuse, resulting in potential copyright violations by fine-tuned LLMs. Therefore, we propose an evaluation benchmark to estimate the risk of Verilog-trained LLMs to generate copyright-protected codes. To minimize this risk, we present an open-source Verilog dataset, FreeSet, containing over 220k files, along with the automated dataset curation framework utilized to provide additional guarantees of fair-use Verilog data. We then execute an LLM fine-tuning framework consisting of continual pre-training, resulting in a fine-tuned Llama model for Verilog, FreeV. Our results indicate that FreeV demonstrates the smallest risk of copyright-infringement among prior works, with only a 3% violation rate. Furthermore, experimental results demonstrate improvements in Verilog generation functionality over its baseline model, improving VerilogEval pass@10 rates by over 10%.
format Preprint
id arxiv_https___arxiv_org_abs_2505_06096
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Free and Fair Hardware: A Pathway to Copyright Infringement-Free Verilog Generation using LLMs
Bush, Sam
DeLorenzo, Matthew
Tieu, Phat
Rajendran, Jeyavijayan
Artificial Intelligence
Limitations in Large Language Model (LLM) capabilities for hardware design tasks, such as generating functional Verilog codes, have motivated various fine-tuning optimizations utilizing curated hardware datasets from open-source repositories. However, these datasets remain limited in size and contain minimal checks on licensing for reuse, resulting in potential copyright violations by fine-tuned LLMs. Therefore, we propose an evaluation benchmark to estimate the risk of Verilog-trained LLMs to generate copyright-protected codes. To minimize this risk, we present an open-source Verilog dataset, FreeSet, containing over 220k files, along with the automated dataset curation framework utilized to provide additional guarantees of fair-use Verilog data. We then execute an LLM fine-tuning framework consisting of continual pre-training, resulting in a fine-tuned Llama model for Verilog, FreeV. Our results indicate that FreeV demonstrates the smallest risk of copyright-infringement among prior works, with only a 3% violation rate. Furthermore, experimental results demonstrate improvements in Verilog generation functionality over its baseline model, improving VerilogEval pass@10 rates by over 10%.
title Free and Fair Hardware: A Pathway to Copyright Infringement-Free Verilog Generation using LLMs
topic Artificial Intelligence
url https://arxiv.org/abs/2505.06096