## AI ESG Classification

On this page, I share ESG classifications of shareholder proposals generated by an AI model (Llama 3.3 70B Instruct), using the methodology developed in my paper: AI Meets ESG: Systematic ESG Classification of Shareholder Proposals. The classifications are constructed for the ISS Shareholder Proposals dataset from 2006 to 2021 and rely primarily on textual inputs from publicly available proxy statements filed with the SEC EDGAR. Additional methodological details are provided in Section 2 of the paper.

The dataset has the following structure:

• ItemOnAgendaID_Final: The primary identifier for shareholder proposals. The original ISS Shareholder Proposals dataset contains some inconsistencies in ItemOnAgendaID relative to other ISS data components in Voting Analytics. These issues were corrected through manual review, which is reflected in the “_Final” suffix, and ItemOnAgendaID_Final can be used as the merge key for ItemOnAgendaID in all ISS Voting Analytics data components. 

• AI_ESG: ESG classification assigned to each shareholder proposal by the AI model

  • E: Environmental

  • S: Social

  • G: Governance

Please feel free to use this dataset in your research. Download here!

I obtain the ESG classifications using the transformers library (Hugging Face) version 4.57.1 and Meta’s Llama 3.3 70B Instruct model (commit 6f6073b423013f6a7d4d9f39144961bfbfbc386b), with PyTorch 2.9.0 + CUDA 13.0 in Python 3.14.0. I use a particular combination of parameters for the model so that each time the same prompt is submitted under identical conditions, the model produces the exact same classification. That being said, ESG classifications may vary if alternative model architectures, parameterizations, or prompt formulations are employed. The current version was compiled on 2025-12-16.

Feel free to contact me if you have any questions, suggestions, ideas regarding the dataset.