cmboulanger/nuextract3-toc

Overview

cmboulanger/nuextract3-toc is a 4.5 billion parameter vision-language model, built upon the Qwen3.5 architecture and fine-tuned using LoRA. Its primary function is to extract structured table-of-contents (TOC) entries from scanned images of TOC pages. The model identifies key information such as title, authors, printed_page_number, and a skip boolean, which indicates if an entry is front/back matter rather than a main chapter.

Key Capabilities

• Structured TOC Extraction: Parses image inputs to output JSON containing detailed TOC entries.
• Precision Improvement: Fine-tuning significantly boosts precision in entry detection, reducing over-generation compared to the base numind/NuExtract3 model.
• Specialized Domain: Optimized for German-language academic book TOCs, trained on data from the Deutsche Nationalbibliothek (DNB).
• Direct Replacement: The adapter is merged, allowing it to be used as a drop-in replacement for numind/NuExtract3 without PEFT dependencies.

Performance

Evaluated on a 44-book held-out split, the model demonstrates notable improvements:

• Precision: Increased from 0.6131 (baseline) to 0.7555 (+0.142).
• F1 Score: Improved from 0.6783 (baseline) to 0.7544 (+0.076).

This gain is primarily due to better precision, as the fine-tuned model generates a more accurate count of entries, aligning with expected ground truth.

Training Data

The model was trained on 624 books (15,372 TOC entries) from the dnb-toc-ground-truth pilot corpus, consisting of DNB table-of-contents scans under a CC0-1.0 license.

When to Use This Model

This model is specifically designed for:

• Automated processing of digitized academic books, particularly those in German.
• Extracting structured metadata from scanned table-of-contents pages (1-3 pages, up to 1200px on the longer side).

Expect degraded performance outside of its intended domain (e.g., other languages, non-academic documents, or non-TOC pages).