hiennthp/esg-bank-model-v3

ESG Greenwashing Detection Model

Multi-task PhoBERT model for Vietnamese ESG content analysis.

Model Architecture

4-task learning:

1. Greenwashing Classification (Legitimate/Greenwashing/Uncertain)
2. ESG Pillar Classification (Environmental/Social/Governance/General)
3. Content Quality Scoring (0-100)
4. ESG Score Prediction (0-100)

Training

• Base Model: vinai/phobert-base
• Strategy: stratified_group_kfold_5
• Folds: 5
• Total Samples: 5214

Performance

Greenwashing Detection

• F1 Score: 0.596
• Precision: 0.624
• Recall: 0.605

Pillar Classification

• Accuracy: 0.411
• F1 Macro: 0.403

Quality Scoring

• MAE: 8.326
• R²: 0.401

ESG Score Prediction

• MAE: 10.318
• R²: 0.405
• Correlation: 0.659

Usage

from transformers import AutoTokenizer, AutoModel
import torch

tokenizer = AutoTokenizer.from_pretrained("hiennthp/esg-bank-model-v3")
# Load model architecture then weights
# model = MultiTaskPhoBERT(config)
# model.load_state_dict(torch.load("best_model_fold0.pt"))

Files

• best_model_fold0.pt - Fold 0 model weights
• best_model_fold1.pt - Fold 1 model weights
• best_model_fold2.pt - Fold 2 model weights
• best_model_fold3.pt - Fold 3 model weights
• best_model_fold4.pt - Fold 4 model weights
• step5_metrics.json - Detailed metrics with per-fold breakdown
• tokenizer/ - PhoBERT tokenizer files

Citation

@software{esg_greenwashing_model,
  author = {ESG Research Team},
  title = {Vietnamese ESG Greenwashing Detection Model},
  year = {2026},
  url = {https://huggingface.co/hiennthp/esg-bank-model-v3}
}