Work

Code and data for: Reading Sustainability Reports as Pages, Not Text: visual retrieval and VLM extraction of GHG metrics

Year: 2026

Type: software

Abstract: Code and outputs of the experiment behind the article "Reading Sustainability Reports as Pages, Not Text: visual retrieval and VLM extraction of GHG metrics". Report pages are kept as images, retrieved with ColQwen2.5 multi-vector embeddings stored in Qdrant, and read by a vision-language model that returns Scope 1, 2 (market- and location-based) and 3 emissions for 2013-2022 as structured JSON. The repository includes the ingestion and extraction pipelines, the extractions of the reported run, and the notebook that evaluates them cell by cell against the GIST gold standard (Beck et al., 2025).

Source: Zenodo (CERN European Organization for Nuclear Research)

DOI: 10.5281/zenodo.22942235

Authors: Lucas, Afonso, Oliveira, Julianne, Silva, Roberto

Institutions: Universidade do Estado do Amazonas, Universidade Federal do Amazonas

Language: English

Topic: Topic Modeling

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

SDG: Climate action

Open Access status: green

OpenAlex ID: W7214166345

Cites:

Cited by:

Related to: 10