TDWG2026
Home / Talks / SAL B
talk · Thursday 24 September · 11:27–⁠11:39 · SAL B

AI Ready Standards with Croissant for Type Specimens Catalog Datasets

Sefika Efeoglu · AI-Readiness Metrics and Metadata for Biodiversity Data

Recording time 4:19:54–4:29:24Open on Vimeo ↗

Summaries and key moments are machine-produced. Check the recording for the speaker’s exact words.

The short versionCombining LLM extraction, Darwin Core and Croissant metadata offers a route from historical specimen catalogues to FAIR, ML-ready datasets.

Overview

What this was about

Sefika Efeoglu presented a workflow for turning historical type specimen catalogues into AI-ready datasets: LLMs (e.g. GPT and Qwen) extract structured information from unstructured catalogue text, which is expressed in Darwin Core and then described with ML Commons Croissant metadata in JSON-LD. She showed how Darwin Core fields map to Croissant fields (scientific name as label, genus as taxonomic feature, individual count as integer, catalogue number as identifier, locality, event date, collector, type status) and how the pipeline evaluates FAIR metrics. The approach aims to bridge biodiversity standards and ML dataset platforms such as Hugging Face, Kaggle and OpenML.

Why it matters. Croissant is becoming the metadata layer ML platforms understand; mapping Darwin Core into it could make collection data directly discoverable and loadable in ML ecosystems.

Key ideas

In the room

  • Biodiversity data are used for species identification, named entity recognition, taxonomic knowledge extraction and knowledge graphs, but historical catalogues are largely unstructured text.
  • FAIRness supports ML reproducibility, clarity, sharing and more reliable model development; the aim is to extend FAIR into ML workflows.
  • Croissant (ML Commons) is a standardised, machine-readable metadata format for ML-ready datasets supported by platforms such as Hugging Face, Kaggle and OpenML: 'create once, publish anywhere, use everywhere'.
  • Workflow: type specimen catalogue, unstructured text, LLM-assisted extraction (GPT, Qwen), Darwin Core metadata, Croissant metadata, AI-ready dataset.
  • Mapping examples: scientificName as label for classification, genus as taxonomic feature, individualCount as integer feature, catalogNumber as unique identifier, locality, eventDate, collector (provenance) and typeStatus.
  • Records are represented in Croissant JSON-LD with explicit field data types; the pipeline measures FAIR metrics and metadata quality.
Jump in

Notable moments

In their words

Transcript

Automatically generated captions can contain mistakes, especially in names and technical terms. Times are relative to the room recording.

Read the transcript ↓
Loading transcript…
Keep exploring

Related talks.

See the map ↗
talk · 22 Sept · 11:15–⁠11:30 · SAL B

Machine-Ready Botanical Data: Mapping Integrated Collections and Traits with Darwin Core Data Package

Darwin Core Data Package can serve as a semantic integration layer for a botanical garden's diverse collections and trait data, and MBG's draft mapping is offered for community refinement.

Nathan VanderKraats
talk · 25 Sept · 09:48–⁠10:00 · SAL A

Extracting Structured Biodiversity Knowledge from Literature with AI-Assisted Workflows

A hybrid text-layer/OCR tool can free tables from grey-literature PDFs into CSV with about 83% precision, but it misses a third of tables, especially narrow ones.

Yağmur Güleç
talk · 22 Sept · 14:40–⁠14:55 · SAL A

BHL and the Planetary Knowledge Base: Structuring Biodiversity Literature for the Next Generation

NHM is re-OCRing all of BHL with modern models and extracting entities into a knowledge graph, with confidence gating to protect BHL data quality.

Qianqian Hiris Gu
Programme abstract

AI Ready Standards with Croissant for Type Specimens Catalog Datasets

Sefika Efeoglu

The FAIR principles are fundamental to Open Science. Publishing datasets used in Artificial Intelligence (AI) experiments in accordance with these principles is essential for ensuring reproducibility. Additionally, these FAIR datasets enhance the performance of Machine Learning (ML) applications, e.g., species labeling using named entity recognition models. Recently, ML Commons introduced Croissant, an ML-ready dataset format designed for seamless integration with ML models and with ML-ready dataset-sharing platforms such as Hugging Face.

Biodiversity data mobilization necessitates adherence to FAIR principles that go beyond merely sharing data in structured formats. We present an experimental approach to converting unstructured text from entomological type specimen catalogs into Darwin Core (DwC) metadata using Large Language Models, including GPT and Qwen. Subsequently, we map Croissant metadata fields to DwC terms-such as dwc:genus, dwc:scientificName, and dwc:individualCount-because these terms can be used to make type specimen catalogs machine-readable.

To evaluate whether our datasets align with FAIR principles for ML, we generate and conduct experiments on subsets of them. These metadata are then integrated into the NFDI4DataScience metadata portal, where FAIR metrics are visualized using the Piveau metric service. This portal provides metadata on datasets for AI use and metrics on FAIR principles. Complementing this quantitative assessment, we validate our metadata format using the FAIR Checker service.

Our proposed metadata framework for type specimen catalogs serves as a reference model for broader applications within biodiversity research domains. In addition, we provide a standardized protocol for the open mobilization of entomological type specimen data.

Submitted to the conference programme before the talk, so the talk itself may differ.