LoRA adapter to correct operational KPI sentiment bias in FinBERT

#1
by yahoyaho13 - opened

Hi everyone 👋

I’m sharing a small but focused LoRA adapter for ProsusAI/finbert that addresses a common issue in financial sentiment models:

Operational KPI statements where decreases imply improvement are often misclassified as negative.

For example:

  • “Failure rate down 10% QoQ”
    → Base FinBERT: Negative
    → +LoRA Adapter: Positive

The adapter is trained on a small, domain-specific dataset and is not intended as a general-purpose replacement, but rather as a targeted correction for manufacturing, operations, and enterprise KPI contexts.

Key points

  • Corrects sentiment for decreasing negative KPIs (defect rate, failure rate, etc.)
  • Preserves negative sentiment for genuine financial deterioration
  • Implemented via LoRA (PEFT) on top of ProsusAI/finbert
  • Includes a Base vs Adapter comparison in the model card

Feedback and suggestions are very welcome. Thanks!

Sign up or log in to comment