Source-Grounded Data Generation for Text-to-JSON Learning

2026-06-18 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

STAGE (Spreadsheet-grounded Text-to-JSON Artifact GEneration) is a novel data generation pipeline designed to address the challenge of creating reliable and scalable text-to-JSON training data for extracting structured information from long, unstructured documents like financial filings or clinical records. This pipeline utilizes Large Language Models (LLMs) for the scalable synthesis of reports and JSON schemas, critically validating the generated ground-truth values against an underlying spreadsheet. Evaluations on the STAGE-Eval benchmark, which includes an 851-example test set, demonstrate that STAGE produces superior training data compared to existing methods. This improvement significantly boosts the Qwen3-4B model's exact match performance from 31.37% to 74.27% and its value accuracy from 45.46% to 90.69%.

Key takeaway

For Machine Learning Engineers developing text-to-JSON extraction systems, STAGE offers a robust method to overcome training data scarcity and quality issues. You should consider integrating source-grounded data generation pipelines, like STAGE, to synthesize high-fidelity training examples. This approach can significantly improve model performance, as demonstrated by the Qwen3-4B's exact match increase to 74.27%, making your extraction solutions more reliable for complex, unstructured documents.

Key insights

STAGE generates high-quality, validated text-to-JSON training data using LLMs and spreadsheets, significantly improving extraction accuracy.

Principles

• LLMs can synthesize structured data.
• Source-grounded validation ensures data quality.
• Better training data boosts model accuracy.

Method

STAGE employs LLMs to synthesize reports and JSON schemas, subsequently validating the generated ground-truth values against an underlying spreadsheet to ensure data integrity.

In practice

• Structure financial filing data.
• Automate clinical record extraction.
• Enhance text-to-JSON model training.

Topics

• Text-to-JSON
• Data Generation
• Large Language Models
• Information Extraction
• Structured Data
• Training Data Quality

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Stop Wasting Tokens: A Smarter Alternative to JSON for LLM Pipelines — TOON reduces LLM token waste by compactly representing repetitive structured data, optimizing prompt efficiency.
• Part 2: Turning My Writing into a Training Set — A method called "grounded generation" creates personalized fine-tuning datasets by constraining LLMs to source documents, ensuring authentic voice and content.
• ExtractBench: A Benchmark and Evaluation Methodology for Complex Structured Extraction — LLMs struggle with complex, enterprise-scale PDF-to-JSON extraction, especially as schema breadth increases.
• Get Vision LLMs to Follow Your Rules: Prompt-Guided JSON Formatting — Prompt-guided JSON formatting enables LLMs to generate structured data with specific, user-defined output formats.
• ToolGrad: Efficient Tool-use Dataset Generation with Textual "Gradients" — ToolGrad inverts tool-use dataset generation, creating valid API chains first, then user queries, ensuring 100% success and lower cost.
• Code Is More Than Text: Uncertainty Estimation for Code Generation — Code generation uncertainty estimation requires specialized methods accounting for code's unique properties, outperforming natural language-derived baselines.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.