Skip to main content
completed production-ai

Fraud LLM Finetune

QLoRA fine-tuning pipeline for 3-class fraud narrative classification with a live Cloud Run interactive demo.

Started: March 28, 2026
Completed: April 5, 2026
GitHub Repository Live Demo

Evidence & Verification

Metrics and claims on this page are tied to the linked artifacts below (repository docs, experiment outputs, and deployment pages when available).

1
Cloud Run Services
3
Classes
6/6
P0 Checks
00009
Latest Revision

Tags

llmqlorafraud-detectiontransformersgradio

Technologies

Python Transformers QLoRA Gradio FastAPI

Overview

Fraud LLM Finetune is an end-to-end narrative classification stack for fraud text analysis. It covers data preparation, QLoRA training, evaluation, adapter merge, and deployment to an interactive classifier UI.

Target classes:

  • LEGITIMATE
  • SUSPICIOUS
  • FRAUDULENT

Live Deployment

This project currently exposes a public demo UI and does not publish a separate public API service endpoint.

Pipeline Scope

  • Dataset preparation and JSONL formatting
  • QLoRA training scripts (real and smoke configurations)
  • Evaluation output (accuracy/F1/latency)
  • Adapter merge/export flow
  • FastAPI + Gradio inference app packaging

Validation Snapshot

Phase 3 automated checks passed 6/6 model behavior scenarios:

  • expected classification outcomes for seeded examples,
  • reasoning-rich responses,
  • custom input classification flow.

GitHub Status