gaurav0506/brsr-greenwashing-roberta-base

BRSR Greenwashing Detection Model (RoBERTa-base)

A fine-tuned RoBERTa-base model for binary classification of greenwashing language in corporate sustainability disclosures.

Model Description

This model is a fine-tuned version of roberta-base for detecting greenwashing in Business Responsibility and Sustainability Reports (BRSR). The model classifies text snippets as either GREENWASHING or NOT_GREENWASHING.

Intended Use

This model is released as a base model for further research on greenwashing detection in ESG (Environmental, Social, and Governance) disclosures. It is intended for:

• Academic research on sustainability language and corporate disclosures
• Fine-tuning with domain-specific or human-annotated datasets
• Exploratory analysis of greenwashing patterns in sustainability reports
• Baseline comparisons in NLP research on ESG and climate-related text

Out-of-Scope Use

This model is not intended for:

• Production deployment without additional validation
• Regulatory or legal determinations of greenwashing
• High-stakes decision-making without human oversight

Performance Metrics

The following metrics were obtained on a held-out validation set of 9,825 samples:

Metric	Value
Accuracy	95.67%
Precision	94.52%
Recall	96.42%
F1 Score	0.9546

Important Caveats on Reported Accuracy

⚠️ The reported accuracy of 95.67% represents an upper bound achieved on held-out validation data.

The training and validation data were derived from a semi-automated labeling pipeline that utilized LLM-generated annotations with human adjudication for a subset of samples. As such:

1. Validation accuracy may be optimistic due to potential label noise and distribution similarity between training and validation splits.

2. Real-world performance expectation: When applied to novel, unseen documents with human-adjudicated ground truth, accuracy is expected to range between 65–80%, depending on domain characteristics and annotation quality.

3. Recommended approach: This model should be treated as a pre-trained base for further fine-tuning with high-quality, human-annotated data specific to the target use case.

Training Details

Training Data

The model was fine-tuned on a curated dataset of 55,684 labeled text snippets extracted from Indian BRSR filings. The labeling methodology employed a priority-based resolution strategy:

• ChatGPT 1000: 1,000 samples with human-verified confidence scores (highest priority)
• Adjudicated Labels: 962 human-adjudicated edge cases
• LLM-Labeled (GPT-4): ~36,000 programmatically labeled samples
• Validation Labels: ~24,000 programmatic labels

Training Configuration

• Base Model: roberta-base (125M parameters)
• Epochs: 6
• Batch Size: 16 (effective: 32 with gradient accumulation)
• Learning Rate: 2e-5 with cosine schedule
• Max Sequence Length: 256 tokens
• Loss Function: Weighted cross-entropy (class-balanced)
• Precision: Mixed precision (FP16)

Usage

Loading the Model

from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch

model_id = "gaurav0506/brsr-greenwashing-roberta-base"

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

# Example inference
text = "We are committed to achieving carbon neutrality by 2050."

inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=256)
with torch.no_grad():
    outputs = model(**inputs)
    predictions = torch.softmax(outputs.logits, dim=-1)

labels = ["NOT_GREENWASHING", "GREENWASHING"]
predicted_label = labels[predictions.argmax().item()]
confidence = predictions.max().item()

print(f"Prediction: {predicted_label} (confidence: {confidence:.2%})")

Using with Pipeline API

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="gaurav0506/brsr-greenwashing-roberta-base"
)

result = classifier("Our sustainability initiatives have transformed our operations.")
print(result)

Greenwashing Categories

The model was trained to detect the following categories of greenwashing, based on regulatory frameworks and academic literature:

1. Vague Claims: Broad, unsubstantiated environmental statements
2. Selective Disclosure: Highlighting positive metrics while omitting negative impacts
3. False Certifications: Claiming non-existent or misleading certifications
4. Irrelevant Claims: Emphasizing minor efforts to distract from larger issues
5. Hidden Trade-offs: Focusing on one attribute while ignoring environmental costs
6. No Proof: Claims without supporting evidence or data

Limitations

• Domain Specificity: Trained primarily on Indian BRSR filings; performance may vary on other ESG frameworks (GRI, SASB, TCFD)
• Language: English only
• Label Quality: Training labels derived from LLM annotations may contain noise
• Temporal Scope: Trained on reports from a specific time period; sustainability language evolves
• Context Length: Maximum 256 tokens; longer passages require chunking

Ethical Considerations

• Greenwashing detection involves subjective judgments that may differ across stakeholders
• False positives could unfairly penalize legitimate sustainability efforts
• False negatives could allow misleading claims to go undetected
• This model should augment, not replace, human expert review

Citation

If you use this model in academic research, please cite:

@misc{greenwashing-roberta-2024,
  author = {Gaurav},
  title = {BRSR Greenwashing Detection Model (RoBERTa-base)},
  year = {2024},
  publisher = {Hugging Face},
  url = {https://huggingface.co/gaurav0506/brsr-greenwashing-roberta-base}
}

Related Resources

• GitHub Repository: Greenwashing-Detection-in-ESG-Reports
• Base Model: roberta-base

License

This model is released under the MIT License.