BERT for Startup SDG Classification

This is a bert-base-uncased model fine-tuned to classify startup company descriptions into one of the 17 UN Sustainable Development Goals (SDGs), plus a "no-impact" category.
This model was trained by Kfir Bar as part of the research paper: "Using Language Models for Classifying Startups Into the UNโ€™s 17 Sustainable Development Goals" (2022).
This repository is hosted by Alon Mannor to make the original model weights accessible to the public.

Model Details

  • Base Model: bert-base-uncased
  • Task: Text Classification
  • Labels: 18 (0: No Impact, 1-17: corresponding SDG)

Label Mapping (id2label)

The model outputs a logit for each of the 18 classes.
The mapping from the index (ID) to the label name is as follows:

{
  "0": "0: No Impact",  
  "1": "SDG 1: No Poverty",  
  "2": "SDG 2: Zero Hunger",  
  "3": "SDG 3: Good Health and Well-being",  
  "4": "SDG 4: Quality Education",  
  "5": "SDG 5: Gender Equality",  
  "6": "SDG 6: Clean Water and Sanitation",  
  "7": "SDG 7: Affordable and Clean Energy",  
  "8": "SDG 8: Decent Work and Economic Growth",  
  "9": "SDG 9: Industry, Innovation and Infrastructure",  
  "10": "SDG 10: Reduced Inequality",  
  "11": "SDG 11: Sustainable Cities and Communities",  
  "12": "SDG 12: Responsible Consumption and Production",  
  "13": "SDG 13: Climate Action",  
  "14": "SDG 14: Life Below Water",  
  "15": "SDG 15: Life on Land",  
  "16": "SDG 16: Peace and Justice Strong Institutions",  
  "17": "SDG 17: Partnerships to achieve the Goal"  
}

How to Use:

You can use this model directly with the text-classification pipeline.

from transformers import pipeline

# Load the classifier
classifier = pipeline("text-classification", model="amannor/bert-base-uncased-sdgclassifier")

# Example description
text = "Our company develops innovative, low-cost solar panels to bring electricity to rural communities."

# Get prediction
result = classifier(text)
print(result)
# [{'label': 'SDG 7: Affordable and Clean Energy', 'score': 0.98...}]

# Example of a non-impact startup
text_2 = "We are a B2B platform for optimizing advertising spend on social media."
result_2 = classifier(text_2)
print(result_2)
# [{'label': '0: No Impact', 'score': 0.95...}]

Training Data:

The model was trained on a dataset of 4,247 startup descriptions (from the Gidron et al. 2023 extension) aggregated from two main sources, which were manually annotated by experts:

  1. Rainmaking (Compass): A global database of impact-focused startups.
  2. Start-up Nation Central (SNC): A database of Israeli startups, including both impact and non-impact companies.

Performance

The model was evaluated on a test set of 866 startups from the original paper.

Task F1-Weighted
18-Label (Full) 0.79
6-Label (5Ps) 0.836

The performance for the 6-label task (People, Planet, Prosperity, Peace, Partnerships, No-Impact) was aggregated from the 18-label predictions.

Citation:

If you use this model or its underlying research, please cite the original paper:
@inproceedings{bar2022usinglm,
title={Using Language Models for Classifying Startups Into the UNโ€™s 17 Sustainable Development Goals},
author={Bar, Kfir},
booktitle={Anonymous Submission to IJCAI-22},
year={2022},
url={httpsall://github.com/Amannor/sdg-codebase/blob/master/articles/IJCAI_2022_SDGs_Methodology.pdf} }

Downloads last month
5
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for amannor/bert-base-uncased-sdgclassifier

Finetuned
(7,257)
this model