AI Infrastructure

Open Financial LLM Leaderboard

Specialized zero-shot financial evaluation framework for language models.

The Open Financial LLM Leaderboard (OFLL) provides a transparent, zero-shot benchmarking environment tailored exclusively for models handling complex financial reports, market forecasting, and sentiment metrics.

Overview

General NLP benchmarks do not measure stock prediction capability or credit risk scoring accurately. OFLL covers seven financial categories (Information Extraction, Textual Analysis, QA, Text Generation, Risk, Forecasting, Decision-Making) using real-world datasets.

Problem

  • Traditional AI benchmarks evaluate general conversation but fail on high-dimensional financial logic.
  • Evaluating model accuracy on specialized financial texts requires standardized zero-shot datasets.

Motivation

Developed under the Fintech Open Source Foundation (FINOS) and SecureFinAI Lab at Columbia University to provide an authoritative model index for financial institutions.

Solution

A zero-shot evaluation runner that scores financial domains and saves structured results JSON files directly to a Hugging Face space.

Implementation

Gradio interface, Python evaluation scripts, and structured JSON results parsers.

Engineering Decisions

Requiring explicit model SHA hash verification to prevent data-contamination and lookup cheating.

Evaluating models zero-shot to measure true context comprehension rather than prompt-engineered memorization.

Challenges & Lessons

Processing highly unstructured financial filing formats (Form 10-K/10-Q) into clean question-answer tokens.

Leaderboard Ingestion Pipeline

Flagship system

Hugging Face Model Submit
  ↓ (Validation / SHA Check)
Evaluation Queue (eval-queue)
  ↓ (Zero-shot Runner)
Task Evaluators (IE, TA, QA, TG, RM, FO, DM)
  ↓ (Scoring: Accuracy, MCC, ROUGE)
Structured JSON Results
  ↓
Gradio Leaderboard Interface (OFLL)

Zero-shot prevents test leakage.

Results are publicly auditable.

Components

Functional building blocks

The flagship documentation uses the same component taxonomy across every system.

Zero-Shot Evaluator

Runs financial tasks against model weights.

Result Parser

Aggregates scores (Accuracy, F1, MCC) into JSON templates.

Gradio Frontend

Visualizes rankings, filters, and categories.

Features

Core capabilities

These are product features in the engineering sense, not marketing copy.

7 specialized task categories

Zero-shot evaluation mode

Hugging Face model integration

Update Config

Initialize target evaluation parameters.

from src.about import Task
task = Task("StockForecasting", "F1", "direction", category="Forecasting")

Example output

A governed action should return enough evidence to review without reconstructing the entire runtime path.

Registered task: StockForecasting
Metric set to: F1

Documentation

  • OFLL contribution guidelines

Papers & Consultation Briefs

    Benchmarks

    • Pipeline execution (7B model): <45 mins
    • Result parsing latency: <2.3s

    Roadmap

    The roadmap separates what is already shipped from the next layer of work so the public record stays explicit.

    Shipped

    Gradio UI table · 7 task categories · FINOS open source release

    Next

    Automatic model execution pipeline · Multi-modal document parsing support

    Related Systems

    Connected systems

    Systems in the ecosystem that depend on or complement this flagship project.