The FinBen Team · Finance

FinMA

A financial large language model instruction-tuned on a comprehensive financial instruction dataset spanning multiple financial tasks.

Overview

FinMA is part of the PIXIU financial AI project and is an instruction-tuned LLM designed for broad financial task coverage. It is fine-tuned on a diverse instruction dataset covering financial sentiment analysis, named entity recognition, question answering, stock prediction, and credit scoring. FinMA aims to provide a single model capable of handling the wide spectrum of financial NLP tasks rather than requiring task-specific models.

Parameters

7B / 30B variants

Base Model

LLaMA

Training Data

PIXIU financial instruction dataset (136K samples)

Tasks Covered

Sentiment, NER, QA, prediction, credit scoring

License

Apache 2.0

Capabilities

Multi-task financial NLP

Financial sentiment analysis

Credit risk assessment

Financial question answering

Stock movement prediction

Use Cases

Building unified financial AI assistants that handle diverse queries

Automating credit scoring from financial documents

Providing financial Q&A for research analysts

Performing multi-task financial document analysis

Pros

  • +Covers a broad range of financial NLP tasks in one model
  • +Open-source with comprehensive evaluation benchmark (PIXIU)
  • +Multiple model sizes for different deployment needs
  • +Strong instruction-following capabilities for financial queries

Cons

  • -Jack-of-all-trades may underperform task-specific models
  • -Training data size is limited compared to BloombergGPT
  • -Based on older LLaMA architecture
  • -May require additional fine-tuning for production financial use

Pricing

Free and open-source. Self-hosting required. The 7B variant runs on consumer hardware; the 30B variant requires multi-GPU setup.

Related Models