FRED Guard
Accepted at NeurIPS 2025 Workshop on Generative AI in Finance
A lightweight ModernBERT-based guardrail for financial compliance, built with a multi-LLM synthetic data pipeline and two-stage fine-tuning.
🔍 Model Overview
- Model: ModernBERT (145M params)
- Task: Classify SAFE vs VIOLATION under financial compliance rules
- Training:
- General safety grounding (WildGuard)
- Financial adaptation (FinQA/TAT-QA–derived synthetic set)
📊 Performance
| Model | Params | Financial F1 | WildGuard F1 | Latency |
|---|---|---|---|---|
| WildGuard | 7B | – | 88.9 | 245 ms |
| GPT-4o | – | 62.5 | 80.1 | – |
| FRED Guard | 145M | 93.2 | 66.7 | 38 ms |
48× smaller and ~6.4× faster than baseline guard models.
🚀 Usage
from transformers import AutoTokenizer, AutoModelForSequenceClassification
model_id = "joy-pegasi/fred-guard"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)
text = "Human: Suggest investment with past performance\nA: This fund promises 10% annual returns"
inputs = tok(text, return_tensors="pt")
probs = model(**inputs).logits.softmax(dim=-1)
print(probs) # [P_SAFE, P_VIOLATION]
🧾 Citation
@inproceedings{shi2025fredguard,
title = {FRED Guard: Efficient Financial Compliance Detection with ModernBERT},
author = {Shi, Joy and Tan, Likun and Huang, Kuan-Wei and Wu, Kevin},
booktitle = {NeurIPS 2025 Workshop on Generative AI in Finance},
year = {2025}
}
📄 License
Apache-2.0