Model Card Climate-TwitterBERT-step-1

Overview:

Using Covid-Twitter-BERT-v2 (https://huggingface.co/digitalepidemiologylab/covid-twitter-bert-v2) as the starting model, we continued domain-adaptive pre-training on a corpus of firm tweets between 2007 and 2020. The model was then fine-tuned on the downstream task to classify whether a given tweet is related to climate change topics.

The model provides a label and probability score, indicating whether a given tweet is related to climate change topics (label = 'Climate') or not (label = 'Non-climate').

Performance metrics:

Based on the test set, the model achieves the following results:

• Loss: 0.0632 • F1-weighted: 0.9778
• F1: 0.9148 • Accuracy: 0.9775 • Precision: 0. 8841 • Recall: 0. 9477

Example usage:

from transformers import pipeline, AutoTokenizer, AutoModelForSequenceClassification

task_name = 'text-classification'
model_name = 'Climate-TwitterBERT/ Climate-TwitterBERT-step1'

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

pipe = pipeline(task=task_name, model=model, tokenizer=tokenizer)

tweet = "We are committed to significantly cutting our carbon emissions by 30% before 2030."
result = pipe(tweet)
# The 'result' variable will contain the classification output: 'Climate' or 'Non-climate'.

Citation:

@article{fzz2025climatetwitter,
  title={Responding to Climate Change Crisis: Firms' Tradeoffs},
  author={Fritsch, Felix and Zhang, Qi and Zheng, Xiang},
  journal={Journal of Accounting Research},
  year={2025},
  doi={10.1111/1475-679X.12625}
}

Fritsch, F., Zhang, Q., & Zheng, X. (2025). Responding to Climate Change Crisis: Firms' Tradeoffs. Journal of Accounting Research. https://doi.org/10.1111/1475-679X.12625

Framework versions

• Transformers 4.28.1 • Pytorch 2.0.1+cu118 • Datasets 2.14.1 • Tokenizers 0.13.3