Numind

Structured information extraction stands at the intersection of natural language processing and artificial intelligence, where raw text converges with actionable data. For businesses and organizations handling voluminous textual data, the ability to swiftly convert unstructured documents into structured formats represents a transformative capability. NuMind, a startup founded by experienced machine learning veterans, has developed a specialized toolkit that revolutionizes this field through compact, high-performing AI models. By building upon their background at Wolfram Research and Make.org, and partnering with the Laboratory of Formal Linguistics at the University of Paris Cité, NuMind has created a platform that sets new standards for efficiency and accuracy in information extraction. Through its NuExtract technology, the company demonstrates how sophisticated AI capabilities can be achieved with fewer resources than traditional approaches, offering both local and cloud deployment options that make advanced text processing accessible to organizations of all sizes.

NuMind Overview

NuMind, founded by former Wolfram Research and Make.org executives, has developed a specialized Natural Language Processing (NLP) toolset that builds upon the company's background in machine learning and artificial intelligence. Located in Paris, France, and Cambridge, MA, the company works closely with the Laboratory of Formal Linguistics (LLF) at the University of Paris Cité, a joint research unit under the CNRS.

The startup's platform centers around its NuExtract technology, which delivers highly performant structured extraction capabilities using compact AI models that require fewer resources than competitors. Key to NuMind's approach is its teaching workflow, which enables users to create accurate custom models through an iterative process of correction and active learning. This foundation model development process has been shown to produce custom models with fewer than 100 million parameters that outperform well-prompted GPT-3.5 after 10 examples and match even larger models like GPT-4 after 100 examples.

At the core of NuMind's technology are specialized foundation models created through advanced training methodologies on domain-specific datasets. These models power the company's suite of information extraction capabilities, including named entity recognition, classification, and structured extraction, while maintaining document privacy and efficiency. The company offers both offline processing options for local file handling and cloud deployment solutions to suit different use case requirements. Through its open-source NuExtract model released under the MIT license, NuMind enables developers to implement these sophisticated text processing capabilities in a wide range of applications while maintaining flexibility and independence from proprietary technology providers.

NuExtract Technology

Developed as a specialized Large Language Model (LLM) for structured extraction tasks, NuExtract demonstrates superior performance metrics while requiring significantly fewer computational resources. Built upon an innovative training framework that processes between 0.5B to 7B parameters, these foundation models consistently outperform larger competitors, achieving extraction accuracy levels comparable to or surpassing GPT-4o in zero-shot settings.

The model architecture employs advanced encoder-decoder designs, including T5 and pure decoder LLM configurations, with NuMind's current implementation utilizing Phi-3-mini variants ranging from 0.5B to 3.8B parameters. These compact models achieve their remarkable performance through an iterative refinement process that begins with synthetic dataset generation and progresses through specialized fine-tuning protocols. The resulting models demonstrate robust zero-shot capabilities while maintaining the versatility required for diverse structured extraction applications.

The company's foundation model development process builds upon a comprehensive dataset containing approximately 10k common field names, carefully balanced between generic concepts (dates, contact information, dimensions) and industry-specific terminology. This structured approach enables effective domain adaptation across technical documents, legal contracts, and financial reports, with the system designed to handle arbitrarily long documents while maintaining efficient processing capabilities.

NuExtract's performance extends beyond basic entity recognition, supporting complex structured extraction tasks that generate hierarchical JSON outputs representing document structures. The model architecture incorporates sophisticated template generation techniques that automatically annotate text inputs using modern LLM prompts, followed by fine-tuning stages that optimize both efficiency and accuracy. Through this combination of automated template creation and targeted fine-tuning, NuMind achieves extraction performance improvements of up to 10x compared to previous methodologies, while requiring just 5x the amount of annotated data typically needed for comparable tasks.

Active Learning Process

The company's teaching workflow consists of three key components: NuNER, automatic machine learning optimizations, and active learning strategies. This process enables users to create high-quality custom models with minimal effort, allowing for early performance assessment through real-time monitoring.

Users begin by creating custom models through fine-tuning with annotated examples. The system employs an active learning strategy to select the most informative documents for annotation, optimizing the learning process. The workflow demonstrates how the model learns from corrections, as it successfully improved from 25% to 2% error rates for misclassified entities through iterative refinement.

NuMind's approach enables the creation of custom models with fewer than 100 million parameters that outperform well-prompted GPT-3.5 after 10 examples and match even larger models like GPT-4 after 100 examples. The process requires just 5x the amount of annotated data typically needed for comparable tasks while executing these improvements on a CPU.

The foundation model development process begins with synthetic dataset generation using LLM prompts, followed by specialized fine-tuning protocols. The company focuses on creating compact models optimized for specific applications while maintaining robust zero-shot capabilities. Through precise template generation techniques, NuMind achieves extraction performance improvements of up to 10x compared to previous methodologies while requiring significantly fewer computational resources.

NuMind's Foundation Models

The foundation models developed by NuMind address specific NLP tasks including Named Entity Recognition (NER), classification, and structured extraction. These models are designed to be compact while maintaining high performance, with typical sizes under 1GB and the ability to run on standard CPUs rather than requiring GPU resources. The company employs advanced training methodologies using self-supervised learning on diverse multi-domain corpora, with performance improvements reported across multiple industry-specific applications and languages.

The NER foundation model employs RoBERTa architecture and demonstrates superior data efficiency, achieving F1-scores of 0.65 with just 5 examples per concept compared to the 30 required by previous models. This significant improvement spans multiple datasets including BioNLP2004 and MIT Movie, where performance enhancements range from 2x to 10x data efficiency improvements. The model's development process involves creating vector representations of text through self-supervised learning, followed by training specialized binary logistic regression classifiers for each concept category.

Structured extraction capabilities are implemented through NuExtract models, which perform at similar or higher levels than LLMs 100 times larger while requiring orders of magnitude fewer computational resources. These models support both zero-shot deployment and fine-tuning for specific extraction problems, with the ability to process arbitrarily long documents and generate hierarchical JSON outputs representing document structures. The structured extraction workflow enables the creation of task-specific foundation models through a process of synthetic dataset generation, modern LLM prompting for annotation, and compact model fine-tuning, demonstrating performance improvements of up to 10x compared to previous methodologies while requiring just 5x the amount of annotated data typically needed for comparable tasks.

Deployment Options and Performance

From local file processing to cloud deployment, NuMind provides flexible implementation options while maintaining document privacy and efficiency. At the core of their deployment strategy is the company's foundation model architecture, which enables compact, high-performance information extraction while requiring minimal computational resources.

The company's deployment options span multiple vectors. Local processing capabilities are facilitated through straightforward user interfaces, beginning with the simple "Process File" button for quick document analysis. For more integrated applications, NuMind offers comprehensive Docker integration that allows users to wrap models in customizable APIs. This modular approach enables native operation on standard CPU servers, making extensive GPU resources unnecessary for most deployment scenarios.

The company's API framework provides robust structure for model output, with local model APIs delivering JSON-formatted results that detail predictions, probabilities, and positional data. This standardization facilitates seamless integration across diverse applications, with supported data entry formats including single strings or JSON arrays of strings. The response schema follows a consistent pattern: each prediction includes a label, text span, probability score, start position, and end position, providing comprehensive information for subsequent processing steps.

Deployment independence is a core principle of NuMind's architecture, with the company explicitly positioning its tools as a solution for reducing vendor lock-in. This philosophy extends to the technical implementation, where even the largest modelvariants (7B parameters) maintain an efficient footprint under 1GB. These compact models support both offline processing for small-scale projects and scalable cloud deployment for organizations with higher processing requirements.

The company's deployment strategy demonstrates particular effectiveness in handling multilingual and multi-industry applications, key areas of focus for their foundational models. The NuExtract platform, for example, has been shown to perform structured extraction tasks at similar or higher levels than LLMs 100 times larger while requiring orders of magnitude fewer computational resources. This performance efficiency is maintained across various document types, from technical medical reports to legal contracts, while processing arbitrarily long documents and generating hierarchical JSON structures representing document hierarchies.