SLYM06/fintone-phi3-financial-sentiment

FinTone — Financial Sentiment Classifier

Fine-tuned DistilBERT for financial text sentiment classification. Trained on 7,634 real financial news sentences, achieving F1 0.902 — outperforming GPT-4o-mini (F1 0.658) by +24.4 points.

Model Details

• Base model: distilbert-base-uncased (67M parameters)
• Task: 3-class sentiment classification (positive / negative / neutral)
• Dataset: Twitter Financial News Sentiment (9,543 samples)
• Training: 5 epochs, lr=2e-5, batch size=32, fp16
• Developer: Shun Le Yi Mon (Sheryl)

Performance

Model	Weighted F1	Accuracy
GPT-4o-mini (baseline)	0.658	65%
FinTone-DistilBERT	0.902	90%
Delta	+0.244	+25%

Per-class Results

Class	Precision	Recall	F1
Negative	0.83	0.94	0.88
Neutral	0.97	0.90	0.93
Positive	0.74	0.88	0.80

Usage

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="SLYM06/fintone-distilbert-financial-sentiment"
)

result = classifier("The company reported record profits this quarter.")
print(result)
# [{'label': 'positive', 'score': 0.94}]

Key Finding

A domain-adapted 67M parameter model outperforms GPT-4o-mini on financial sentiment by 24.4 F1 points — demonstrating that fine-tuning a purpose-built classifier on domain-specific data is more effective and cost-efficient than prompting a general LLM.

Training Data

Twitter Financial News Sentiment dataset — 9,543 financial news headlines and tweets labelled as positive, negative, or neutral.

GitHub

github.com/Shun024/fintone