# numind-nuextract-foundation-model-structured-extraction-2024-06-24

## Veille

NuExtract NuMind - foundation model for structured JSON extraction, compact format

## Titre Article

NuExtract: A Foundation Model for Structured Extraction

## Date

2024-06-24

## URL

https://numind.ai/blog/nuextract-a-foundation-model-for-structured-extraction

## Keywords

structured extraction, JSON, language model, NLP, fine-tuning, compact LLM, open source, MIT, zero-shot

## Authors

Alexandre Constantin, Liam Cripwell, Etienne Bernard

## Ton

Profile: Technical corporate blog article, popularized scientific register, applied research perspective.
Style: Enthusiastic yet rigorous tone, backed by precise comparative metrics. Uses striking numerical comparisons (100x smaller, 35x more efficient) to demonstrate the value proposition. Authority built on methodology and benchmarks. Target audience: ML practitioners, NLP developers, technical decision-makers evaluating extraction solutions.

## Pense-betes

- Models from 0.5B to 7B parameters rivaling LLMs 100x larger
- NuExtract-tiny outperforms GPT-3.5 while being 100x smaller
- NuExtract-large achieves GPT-4o equivalence with 100x fewer parameters
- Training dataset: 50,000 examples generated from C4 with Llama 3 70B
- Extraction tree depth up to 9 hierarchical levels
- MIT license - open source
- Operates in zero-shot mode or with specialized fine-tuning
- Use cases: medical records, legal documents, financial reports
- Advantages: reduced inference costs, local/private deployment possible

## RésuméDe400mots

NuMind introduces NuExtract, a language model specialized in converting text into structured JSON data. The major innovation lies in the creation of compact models (from 0.5 billion to 7 billion parameters) that match or exceed the performance of generic LLMs a hundred times larger.

The model excels at extracting complex information from varied documents, organizing data hierarchically. Typical use cases include the analysis of medical records, legal documents, and financial reports. NuExtract can operate in zero-shot mode for general tasks or be fine-tuned for specific business problems.

The dataset creation methodology is a key element of the success. The team leveraged 300,000 English texts from the C4 dataset, using Llama 3 70B to generate extraction templates and annotations. This process produced 50,000 quality training examples, with extraction trees reaching up to 9 levels of hierarchical depth.

Performance results demonstrate the effectiveness of the approach. NuExtract-tiny outperforms GPT-3.5 despite being 100 times smaller. The standard NuExtract version surpasses Llama3-70B while being 35 times more compact. NuExtract-large achieves equivalence with GPT-4o with a size reduction factor of 100.

These performances open up considerable practical advantages compared to proprietary alternatives. Inference costs are drastically reduced thanks to the compact size of the models. Local or private deployment becomes feasible, meeting the confidentiality requirements of many organizations. The fine-tuning capability makes it easy to adapt the model to specific domains without deep ML expertise.

The model is distributed under the MIT license, allowing free use in commercial and research contexts. This open source approach positions NuExtract as a credible alternative to proprietary extraction APIs, particularly for organizations concerned with controlling their costs and their data.

The article illustrates an important trend: specialized models of modest size can rival generalist giants on targeted tasks, paving the way for more economical and sovereign deployments.

## GrapheDeConnaissance

- NuMind —a_créé→ NuExtract (TECHNOLOGIE, 0.98)
- NuExtract —permet→ extraction structurée JSON (CONCEPT, 0.97)
- NuExtract-tiny —surpasse→ GPT-3.5 (TECHNOLOGIE, 0.95)
- NuExtract-large —converge_avec→ GPT-4o (TECHNOLOGIE, 0.93)
- NuExtract —est_basé_sur→ Llama 3 70B (TECHNOLOGIE, 0.9)
- NuExtract —utilise→ dataset C4 (TECHNOLOGIE, 0.92)
- NuExtract —utilise→ licence MIT (CONCEPT, 0.98)
- Alexandre Constantin —a_créé→ NuExtract (TECHNOLOGIE, 0.95)
- Liam Cripwell —a_créé→ NuExtract (TECHNOLOGIE, 0.95)
- Etienne Bernard —a_créé→ NuExtract (TECHNOLOGIE, 0.95)
- NuExtract —réduit→ coûts d'inférence (CONCEPT, 0.9)
- modèles spécialisés compacts —concurrence→ LLM généralistes (CONCEPT, 0.88)

---
Canonical: https://www.thekb.eu/en/fiches/numind-nuextract-foundation-model-structured-extraction-2024-06-24/