[Paper Review] SciInstruct: a Self-Reflective Instruction Annotated Dataset for Training Scientific Language Models

Dan Zhang, Ziniu Hu+7 more|arXiv (Cornell University)|Jan 15, 2024

Topic Modeling4 citations

TL;DR

This paper introduces SciInstruct, a high-quality, self-reflective instruction-annotated dataset for scientific reasoning, and SciGLM, a suite of fine-tuned scientific language models. Using an LLM-based self-reflective framework to generate and correct reasoning steps, the authors create a diverse dataset spanning math, physics, chemistry, and formal proofs, which enables significant performance gains in scientific and mathematical reasoning without sacrificing base language capabilities.

ABSTRACT

Large Language Models (LLMs) have shown promise in assisting scientific discovery. However, such applications are currently limited by LLMs' deficiencies in understanding intricate scientific concepts, deriving symbolic equations, and solving advanced numerical calculations. To bridge these gaps, we introduce SciInstruct, a suite of scientific instructions for training scientific language models capable of college-level scientific reasoning. Central to our approach is a novel self-reflective instruction annotation framework to address the data scarcity challenge in the science domain. This framework leverages existing LLMs to generate step-by-step reasoning for unlabelled scientific questions, followed by a process of self-reflective critic-and-revise. Applying this framework, we curated a diverse and high-quality dataset encompassing physics, chemistry, math, and formal proofs. We analyze the curated SciInstruct from multiple interesting perspectives (e.g., domain, scale, source, question type, answer length, etc.). To verify the effectiveness of SciInstruct, we fine-tuned different language models with SciInstruct, i.e., ChatGLM3 (6B and 32B), Llama3-8B-Instruct, and Mistral-7B: MetaMath, enhancing their scientific and mathematical reasoning capabilities, without sacrificing the language understanding capabilities of the base model. We release all codes and SciInstruct at https://github.com/THUDM/SciGLM.

Motivation & Objective

• To address the scarcity of high-quality, reasoning-rich scientific instruction data for training science-specific language models.
• To improve the scientific reasoning capabilities of large language models beyond narrow mathematical tasks, including physics, chemistry, and formal proofs.
• To develop a scalable, AI-driven annotation framework that reduces human dependency while maintaining high data quality through self-reflection and error correction.
• To demonstrate that fine-tuning on diverse, high-quality scientific instruction data significantly enhances reasoning performance across multiple benchmarks.
• To release SciInstruct, SciGLM, and the self-reflective framework to support open research in scientific AI.

Proposed method

• Employing a self-reflective instruction annotation framework where LLMs generate reasoning steps and then critique and revise their own outputs based on incorrect answers.
• Using LLMs to generate both positive and negative samples for training a data classifier to filter noisy or low-quality instructions.
• Constructing SciInstruct by combining curated scientific questions from math, physics, chemistry, and formal proofs in Lean, with step-by-step reasoning and verified answers.
• Fine-tuning the ChatGLM3 series models (6B, 12B, 32B) on SciInstruct to create SciGLM, enhancing scientific reasoning while preserving base language understanding.
• Applying a multi-stage data processing pipeline: initial LLM generation, self-reflection on errors, answer verification, and classifier-based filtering for high-quality data.
• Leveraging existing LLMs to autonomously generate and refine reasoning chains, minimizing human annotation while ensuring correctness through iterative self-critique.

Figure 1: Average accuracy on CEval-Sci and SciEval benchmarks of different LLMs. Our SciGLM Models fine-tuned on ChatGLM3 are shown in bold.

Experimental results

Research questions

• RQ1 Can an LLM-based self-reflective framework produce high-quality scientific instruction data with minimal human intervention?
• RQ2 Can fine-tuning on a diverse, reasoning-rich dataset like SciInstruct significantly improve scientific reasoning in LLMs across multiple domains?
• RQ3 Does enhancing reasoning capabilities through such data improve performance on scientific benchmarks without degrading general language understanding?
• RQ4 To what extent can synthetic, self-annotated data match or exceed the performance of human-annotated instruction datasets in scientific reasoning tasks?
• RQ5 Can a self-improving data pipeline using LLMs and classifiers effectively filter out noisy or incorrect instructions in scientific domains?

Key findings

• SciGLM models fine-tuned on SciInstruct achieve state-of-the-art performance on CEval-Sci and SciEval benchmarks, outperforming both the base ChatGLM3-6B-Base and larger models with orders of magnitude more parameters.
• The self-reflective annotation framework successfully generates high-quality reasoning steps with minimal human input, significantly improving data quality and model performance.
• SciGLM maintains strong base language understanding capabilities while achieving substantial gains in scientific reasoning, indicating no trade-off between general and specialized abilities.
• The model shows consistent improvements across all tested sizes (6B, 12B, 32B), demonstrating scalability and generalization of the training approach.
• The data classifier used in the pipeline improves instruction quality, and future bootstrapping of data could further enhance model performance.
• The release of SciInstruct, SciGLM, and the self-reflective framework enables broader research in scientific language modeling and reasoning.

Figure 2: The pipeline of constructing SciInstruct . On the far left is a mix of training datasets. The purpose of the annotation is to supplement chain-of-thought processes with reflective generation. The goal of the filter is to train a data classifier and generate high-quality instructions.