| --- |
| language: |
| - en |
| license: apache-2.0 |
| library_name: transformers |
| tags: |
| - text-classification |
| - bert |
| - finbert |
| - finance |
| - sentiment |
| - sentiment-analysis |
| - financial-sentiment |
| datasets: |
| - FinanceInc/auditor_sentiment |
| - nickmuchi/financial-classification |
| - warwickai/financial_phrasebank_mirror |
| pipeline_tag: text-classification |
| --- |
| |
| # 🎯 FinBERT-Pro |
|
|
| An improved financial sentiment model built on [ProsusAI/finbert](https://huggingface.co/ProsusAI/finbert). Fine-tuned on 3 expert-annotated financial datasets for more robust sentiment classification. |
|
|
| The model provides softmax outputs for three sentiment classes: **Positive**, **Negative**, **Neutral**. |
|
|
| ## 🚀 Usage |
|
|
| ```python |
| from transformers import pipeline |
| |
| classifier = pipeline("text-classification", model="ENTUM-AI/FinBERT-Pro") |
| |
| classifier("Stock price soars on record-breaking earnings report") |
| # [{'label': 'Positive', 'score': 0.99}] |
| |
| classifier("Company announces quarterly earnings results") |
| # [{'label': 'Neutral', 'score': 0.98}] |
| |
| classifier("Revenue decline signals weakening market position") |
| # [{'label': 'Negative', 'score': 0.98}] |
| ``` |
|
|
| ## 📊 Training Data |
|
|
| Fine-tuned on 3 expert-annotated public datasets: |
|
|
| | Dataset | Samples | |
| |---------|---------| |
| | [FinanceInc/auditor_sentiment](https://huggingface.co/datasets/FinanceInc/auditor_sentiment) | ~4.8K | |
| | [nickmuchi/financial-classification](https://huggingface.co/datasets/nickmuchi/financial-classification) | ~5K | |
| | [warwickai/financial_phrasebank_mirror](https://huggingface.co/datasets/warwickai/financial_phrasebank_mirror) | ~4.8K | |
|
|
| Unlike the original FinBERT (trained on a single dataset), FinBERT-Pro combines multiple expert-annotated sources for better generalization across different financial text styles. |
|
|
| ## 🔍 What's Different from FinBERT? |
|
|
| - **Multiple data sources** — trained on 3 expert-annotated datasets instead of 1 |
| - **Class-weighted training** — handles imbalanced label distributions |
| - **Better generalization** — diverse training data improves robustness on unseen financial texts |
|
|
| ## ⚠️ Limitations |
|
|
| - English only |
| - Designed for short financial texts (headlines, news, reports) |
|
|
