Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

NeMo-Inspector: A Visualization Tool for LLM Generation Analysis

Daria Gitman, Igor Gitman, Evelina Bakhturina

TL;DR

NeMo-Inspector addresses the challenge of validating synthetic data for LLM adaptation by providing an integrated toolkit for inference-driven generation analysis. It combines two pages (Inference and Analyze) to support homogeneous and heterogeneous generation assessment, custom Python statistics, and manual annotation. The paper demonstrates significant data-quality improvements on GSM-Plus (reducing low-quality samples from 46.99% to 19.51%) and identifies generation error patterns in OpenMath-Mistral-7B-v0.1, yielding positive accuracy gains on MATH ($\Delta = 1.92\%$) and GSM8K ($\Delta = 4.17\%$) after targeted data improvements. The work underscores the practical impact of interactive data exploration for rapid prompt engineering and synthetic data refinement.

Abstract

Adapting Large Language Models (LLMs) to novel tasks and enhancing their overall capabilities often requires large, high-quality training datasets. Synthetic data, generated at scale, serves a valuable alternative when real-world data is scarce or difficult to obtain. However, ensuring the quality of synthetic datasets is challenging, as developers must manually inspect and refine numerous samples to identify errors and areas for improvement. This process is time-consuming and requires specialized tools. We introduce NeMo-Inspector, an open-source tool designed to simplify the analysis of synthetic datasets with integrated inference capabilities. We demonstrate its effectiveness through two real-world cases. Analysis and cleaning of the synthetically generated GSM-Plus dataset with NeMo-Inspector led to a significant decrease in low-quality samples from 46.99% to 19.51%. The tool also helped identify and correct generation errors in OpenMath models, improving accuracy by 1.92% on the MATH dataset and by 4.17% on the GSM8K dataset for a Meta-Llama-3-8B model fine-tuned on synthetic data generated from Nemotron-4-340B.

NeMo-Inspector: A Visualization Tool for LLM Generation Analysis

TL;DR

NeMo-Inspector addresses the challenge of validating synthetic data for LLM adaptation by providing an integrated toolkit for inference-driven generation analysis. It combines two pages (Inference and Analyze) to support homogeneous and heterogeneous generation assessment, custom Python statistics, and manual annotation. The paper demonstrates significant data-quality improvements on GSM-Plus (reducing low-quality samples from 46.99% to 19.51%) and identifies generation error patterns in OpenMath-Mistral-7B-v0.1, yielding positive accuracy gains on MATH () and GSM8K () after targeted data improvements. The work underscores the practical impact of interactive data exploration for rapid prompt engineering and synthetic data refinement.

Abstract

Adapting Large Language Models (LLMs) to novel tasks and enhancing their overall capabilities often requires large, high-quality training datasets. Synthetic data, generated at scale, serves a valuable alternative when real-world data is scarce or difficult to obtain. However, ensuring the quality of synthetic datasets is challenging, as developers must manually inspect and refine numerous samples to identify errors and areas for improvement. This process is time-consuming and requires specialized tools. We introduce NeMo-Inspector, an open-source tool designed to simplify the analysis of synthetic datasets with integrated inference capabilities. We demonstrate its effectiveness through two real-world cases. Analysis and cleaning of the synthetically generated GSM-Plus dataset with NeMo-Inspector led to a significant decrease in low-quality samples from 46.99% to 19.51%. The tool also helped identify and correct generation errors in OpenMath models, improving accuracy by 1.92% on the MATH dataset and by 4.17% on the GSM8K dataset for a Meta-Llama-3-8B model fine-tuned on synthetic data generated from Nemotron-4-340B.
Paper Structure (22 sections, 2 figures, 2 tables)

This paper contains 22 sections, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Tool overview
  3. Inference page
  4. Analyze page
  5. Usage scenarios
  6. Synthetic data quality check
  7. Setup
  8. Statistics
  9. Sorting
  10. Filtering, labeling and batch editing
  11. Comparison
  12. Annotation of LLM generations
  13. Labeling
  14. Manual editing
  15. Prompt adjustment
  16. ...and 7 more sections

Figures (2)

  • Figure 1: Comprehensive data analysis with the Analyze page of NeMo-Inspector. The tool offers a holistic data view through three complementary perspectives: Intra-Sample Analysis, Comparative Analysis, and Inter-Parameter Analysis.
  • Figure 2: An example of similar solutions with varying correctness illustrates the incorrect expected answer. On the left is a question from the GSM-Plus dataset, and on the right is a question from the GSM8K dataset.