QuantumLLMInstruct: A 500k LLM Instruction-Tuning Dataset with Problem-Solution Pairs for Quantum Computing

TL;DR

QuantumLLMInstruct introduces a vast, diverse dataset of over 500,000 instruction-following problem-solution pairs tailored for quantum computing, aiming to enhance LLM capabilities in this complex domain.

Contribution

The paper presents the creation of the largest and most comprehensive quantum computing dataset for instruction fine-tuning, utilizing a rigorous multi-stage methodology including domain-specific problem generation and self-assessment.

Findings

Dataset contains over 500,000 high-quality problem-solution pairs.

Incorporates advanced reasoning techniques like Chain-of-Thought and ToRA.

Validated for accuracy and diversity through self-assessment.

Abstract

We present QuantumLLMInstruct (QLMMI), an innovative dataset featuring over 500,000 meticulously curated instruction-following problem-solution pairs designed specifically for quantum computing - the largest and most comprehensive dataset of its kind. Originating from over 90 primary seed domains and encompassing hundreds of subdomains autonomously generated by LLMs, QLMMI marks a transformative step in the diversity and richness of quantum computing datasets. Designed for instruction fine-tuning, QLMMI seeks to significantly improve LLM performance in addressing complex quantum computing challenges across a wide range of quantum physics topics. While Large Language Models (LLMs) have propelled advancements in computational science with datasets like Omni-MATH and OpenMathInstruct, these primarily target Olympiad-level mathematics, leaving quantum computing largely unexplored. The creation of QLMMI follows a rigorous four-stage methodology. Initially, foundational problems are developed using predefined templates, focusing on critical areas such as synthetic Hamiltonians, QASM code generation, Jordan-Wigner transformations, and Trotter-Suzuki quantum circuit decompositions. Next, detailed and domain-specific solutions are crafted to ensure accuracy and relevance. In the third stage, the dataset is enriched through advanced reasoning techniques, including Chain-of-Thought (CoT) and Task-Oriented Reasoning and Action (ToRA), which enhance problem-solution diversity while adhering to strict mathematical standards. Lastly, a zero-shot Judge LLM performs self-assessments to validate the dataset's quality and reliability, minimizing human oversight requirements.