FinRED: An Expert-Guided Benchmark Generation and Evaluation Framework for Financial LLM Red-Teaming (IEEE ICDM 2026)

A red-team benchmark generation pipeline for safety evaluation in the financial domain.

---

Supplementary Documentation

Detailed materials referenced in the paper:

• docs/expert_validation.md — Per-question Focus Group Interview (FGI) results from the 12 FSI domain experts (agreement, Likert scores, Cohen's κ / Krippendorff's α).
• docs/judge_rubric.md — Full LLM-as-a-Judge evaluation template, the five rubric dimensions, decision rule, and output JSON schema.
• docs/schemas.md — Threat behavior schema structure (essential/optional elements) with a worked R4.4 example.
• docs/related_work.md — Per-benchmark descriptions of prior financial-domain LLM benchmarks and how FinRED differs.

---

Project Structure

FinRED/
├── main.py                      # Main runner
├── requirements.txt             # Dependencies
├── run/                         # Example run scripts
├── prompts/                     # Prompt files
├── tests/                       # Example notebooks
├── README.md
│
├── src/
│   ├── Step1_build.py           # Scenario generation module
│   ├── Step2_build.py           # Seed prompt generation module
│   ├── __init__.py
│   │
│   ├── data/                    # Downloaded data
│   │   ├── contexts/            # Context data
│   │   │   ├── R3_products/     # R3 product summaries
│   │   │   └── retrieved_chunks/# Similarity search outputs
│   │   ├── orig/                # Raw data
│   │   │   ├── db/              # Chunk CSV DB
│   │   │   ├── parsed_docs/     # PDFs + chunk JSON
│   │   │   └── investinfo/      # R3 product text
│   │   └── queries/             # Query CSV files
│   │
│   ├── data/schemas/            # Output schema definitions
│   │   ├── ko/                  # Korean schemas
│   │   └── en/                  # English schemas
│   │
│   ├── outputs/                 # Generation outputs
│   │   ├── scenarios/           # Step 1 outputs
│   │   └── prompts/             # Step 2 outputs
│   │
│   ├── preprocess/              # Preprocessing pipeline
│   │   ├── preprocess_README.md # Detailed preprocessing guide
│   │   ├── 1_chunking.py
│   │   ├── 2_parsed_to_csv.py
│   │   ├── 3_common_to_csv.py
│   │   ├── 4_product_summarizer.py
│   │   ├── 5_summary_extractor.py
│   │   └── 6_chunk_retriever.py
│   │
│   ├── eval/                    # Evaluation module
│   │   ├── judge_finred.py      # Evaluation script
│   │   ├── dataset/             # Evaluation dataset

---

Environment Setup

0. Download Related Data

Google Drive: https://drive.google.com/drive/u/0/folders/1cfBf419OUDrQQMRKMPLLJqRX97WMxExC Google Drive

Place the downloaded data to match the folder structure above (under src/data).

1. Create a Virtual Environment (Python 3.10)

# Conda
conda create -n finred python=3.10 -y
conda activate finred

# Or venv
python3.10 -m venv finred_env
source finred_env/bin/activate

2. Install Packages

cd /path/to/FinRED
pip install -r requirements.txt

3. Preprocessing Environment (Optional)

If you need preprocessing (PDF chunking, etc.):

# Unstructured
pip install "unstructured[all-docs]"

# System dependencies (Ubuntu/Debian)
sudo apt-get install -y libmagic-dev poppler-utils tesseract-ocr tesseract-ocr-kor libreoffice pandoc

Preprocessing guide: src/preprocess/preprocess_README.md

4. Verify Installation

python --version  # Python 3.10.x
python -c "import torch; print(f'CUDA: {torch.cuda.is_available()}')"

---

Pipeline Overview

[Context + Schema + Query]
         │
         ▼
   ┌─────────────┐
   │ Step1_build │  Scenario generation (OpenAI GPT-4)
   └─────────────┘
         │
         ▼
   [Scenario JSON]
         │
         ▼
   ┌─────────────┐
   │ Step2_build │  Seed prompt generation (Gemini)
   └─────────────┘
         │
         ▼
   [Seed Prompts]
         │
         ▼
   ┌─────────────┐
   │ Evaluation  │  Model response evaluation
   └─────────────┘

---

Module Details

Step1_build.py - Scenario Generation

Generates red-team scenarios based on context and schema.

• Inputs: schema, queries, retrieved context chunks
• Outputs: scenario JSON files in src/outputs/scenarios/{category}/
• Model: OpenAI GPT-4

Step2_build.py - Seed Prompt Generation

Generates seed prompts from scenarios.

• Inputs: scenario JSON from Step 1
• Outputs: prompt JSON + merged CSV in src/outputs/prompts/{category}/
• Model: Google Gemini 2.5 Pro

---

Usage

Quickstart

Data Generation

# Run the full pipeline with the example script
sh run/run_data_generate.sh

Judge

python src/eval/judge_finred.py \
    -i src/eval/dataset/qwen_2.5_test.csv \
    -o qwen_2.5_test_judged \
    -d src/eval/infer_result

Basic

cd /path/to/FinRED

python main.py \
    --step <1|2|all> \
    --category <category> \
    --openai_api_key "sk-..." \
    --gemini_api_key "AIza..."

Parameters

Parameter	Required	Description
--step	Yes	Step to run: 1 (scenario), 2 (prompt), all (sequential)
--category	Yes	Category: R1, R2, R3, R4, R5 or R1_1, etc.
--openai_api_key	Step 1	OpenAI API key
--gemini_api_key	Step 2	Gemini API key
--lang	Optional	Prompt language: ko (default), en
--num_prompts	Optional	Number of prompts (default: 3)
--step1_model	Optional	Model for Step 1 (default: gpt-4.1-2025-04-14)
--step2_model	Optional	Model for Step 2 (default: models/gemini-2.5-pro)
--model_name	Optional	Deprecated alias for --step1_model

Examples

# Full pipeline (R1)
python main.py \
    --step all \
    --category R1 \
    --openai_api_key "sk-proj-..." \
    --gemini_api_key "AIza..."

# Step 1 only
python main.py \
    --step 1 \
    --category R2 \
    --openai_api_key "sk-proj-..."

# Step 2 only (English, 5 prompts)
python main.py \
    --step 2 \
    --category R5 \
    --lang en \
    --num_prompts 5 \
    --gemini_api_key "AIza..."

# Single subcategory
python main.py \
    --step 2 \
    --category R1_1 \
    --num_prompts 3 \
    --gemini_api_key "AIza..."

---

Category Info

Category	Subcategories	Description
R1	R1_1 ~ R1_6	Financial security threats
R2	R2_1 ~ R2_5	Fraud and illegal activities
R3	R3_1 ~ R3_3	Mis-selling
R4	R4_1 ~ R4_5	Market manipulation
R5	R5_1 ~ R5_7	Regulatory violations

---

Evaluation

Evaluate model responses to generated prompts.

Run

cd /path/to/FinRED/src/eval

python src/eval/judge_finred.py \
    -i <input_csv_path> \
    -o <output_basename> \
    -d <output_dir>

Example

python src/eval/judge_finred.py \
    -i ./dataset/qwen_2.5_7b_all_with_responses.csv \
    -o qwen2.5_7b_responses \
    -d ./infer_result/

Input CSV Format

The CSV should include:

• prompt column
• model response column

Output

• infer_result/: evaluation results
• judge_errors/: error logs

---

Output Formats

Scenario JSON (Step 1)

{
  "scenario_id": "R1_1_0001",
  "context": "...",
  "risk_description": "...",
  "attack_vector": "..."
}

Prompt JSON (Step 2)