roberta_finance_sentiment

RoBERTa-Large Fine-Tuned for Financial Sentiment Analysis

This repository contains a RoBERTa-based model for financial sentiment classification. The model predicts whether a financial news headline or sentence is positive, neutral, or negative.

Model Overview

• Base model: RoBERTa-Large
• Task: Financial sentiment classification (3 classes)
• Training data: Financial news headlines and sentences
• Dataset source: Kaggle - Finance News Sentiments
• Output labels:
  • 0: Negative
  • 1: Neutral
  • 2: Positive

Evaluation Results

• Test Accuracy: 0.7627
• Multiclass ROC AUC (macro-average): 0.9124

Model Folder Structure

roberta_finance_sentiment/
    config.json
    merges.txt
    model.safetensors
    special_tokens_map.json
    tokenizer_config.json
    tokenizer.json
    vocab.json

Note: Only the model files are stored in roberta_finance_sentiment/. Scripts and datasets are kept separate and are not included in this folder or in the model upload.

How to Use the Fine-Tuned Model

1. Load and Use the Model in Python

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification
# Directory of the model folder
model_dir = "roberta_finance_sentiment"
# read the model
tokenizer = AutoTokenizer.from_pretrained(model_dir)
model = AutoModelForSequenceClassification.from_pretrained(model_dir)
model.eval()

# Example
text = "Apple stock surges after strong earnings report."
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True, max_length=128)
with torch.no_grad():
    logits = model(**inputs).logits
    pred = torch.argmax(logits, dim=1).item()

label_map = {0: 'negative', 1: 'neutral', 2: 'positive'}
print(f"Predicted sentiment: {label_map[pred]}")

Notes

• The model was trained and evaluated on data from the Kaggle dataset linked above.
• The roberta_finance_sentiment/ folder contains only the files needed for inference.
• Scripts and datasets are not included in the model folder or in the model upload.
• For best results, use a GPU for inference if available.

Limitations

• Model is trained on headline-level sentiment.
• Sarcasm, irony, or complex phrasing may reduce prediction accuracy.

---

Date: June 2025