cometadata/etd-metadata-parsing-qwen3.5-4b-dapo-v2

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 26, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The cometadata/etd-metadata-parsing-qwen3.5-4b-dapo-v2 model, developed by cometadata, is a 4.5 billion parameter Qwen3.5-based language model specifically fine-tuned for extracting 10-field bibliographic metadata from the front matter of electronic theses and dissertations (ETDs). It processes text from PDF pages and outputs a single JSON record, excelling in structured data extraction with an aggregate score of 0.867 on its test split, outperforming Claude Opus 5 and previous iterations. This model is optimized for precise, multilingual metadata parsing without translation, making it ideal for academic data management.

Loading preview...

Overview

The cometadata/etd-metadata-parsing-qwen3.5-4b-dapo-v2 is a specialized 4.5 billion parameter Qwen3.5-based model designed for highly accurate bibliographic metadata extraction from the front matter of electronic theses and dissertations (ETDs). It takes the text from the initial pages of a thesis and outputs a structured 10-field JSON record, including title, creators, advisors, degree_grantor, and date_issued. This model is a weight-space average of three checkpoints, trained on a 2,999-record dataset with an institution-held-out split, ensuring robust performance on unseen data.

Key Capabilities

  • Precise Metadata Extraction: Achieves an aggregate score of 0.867 and an exact record score of 0.146 on its test split, significantly outperforming Qwen3.5-4B zero-shot (+7.3 points) and Claude Opus 5 3-shot (+3.6 points).
  • Multilingual Support: Extracts metadata in the document's original language across Portuguese, English, Spanish, French, German, and Dutch, without translation.
  • Structured JSON Output: Directly produces a 10-field JSON record, adhering to a strict schema, with thinking disabled to ensure direct JSON responses.
  • Robust Training: Benefits from DAPO (Denoising Autoencoder Pre-training Objective) fine-tuning on a large, curated dataset, including a noise-masked reward mechanism for improved accuracy on challenging fields.

Good for

  • Academic Data Management: Automating the extraction of critical bibliographic information from large collections of ETDs.
  • Library and Archival Systems: Populating metadata fields for digital repositories and catalogs.
  • Research and Analytics: Creating structured datasets from unstructured thesis documents for academic research and trend analysis.

Finetunes
1 models