Skip to content

bedolpab/tuutrag

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

1 Commit
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Teaching to understand, understanding to teach: retrieval-augmented generative reasoning for requirement and test case traceability

Pablo Cesar Bedolla Ortiz, Dominican University

Mentors: Gus Razo and Kae Sawada, Jet Propulsion Laboratory/California Institute of Technology


Abstract

Requirement traceability, validation, and verification can become difficult within engineering projects, notably as they scale. Technical specification documents detailing these structured processes are primarily expressed using natural language. With the adoption of Large Language Models (LLMs) and their effectiveness in natural language processing tasks such as information and relationship extraction, specification documents can be leveraged. While traditional requirement and test engineering methods rely on human thinking, smaller models can perform goal-aligned reasoning when trained with human feedback. We propose an approach for requirement traceability using Light Retrieval-Augmented Generation (RAG) by fine-tuning an LLM for improved knowledge discovery, primarily entity-relationship extraction. Our approach seeks to induce traceability, which emerges from the knowledge graph component of the RAG system by evaluating responses with human expert feedback for alignment. Motivated by the Feynman Technique, we reinforce areas of misunderstanding or misalignment in LLM-based tasks by fine-tuning with Direct Preference Optimization (DPO), which adjusts from response pairs—inadequate (incorrect) and expert-improved (correct). Quantized Low-Rank Adaptation (QLoRA) is employed to optimize fine-tuning under hardware constraints. The implication is linkage between a provided input and a ranked output set as an emergent property of the model’s understanding. Metrics are derived from human cueing and imitation tests.


Memgraph Knowledge Graph

About

Teaching to Understand, Understanding to Teach: Retrieval Augmented Generation for Requirements Traceability

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages