Bachelor's Thesis

Fine-Tuning LLMs for ERP Systems

Optimizing search quality in SAP workflows by fine-tuning the Teuken-7B German LLM to generate structured JSON outputs from unstructured product descriptions.

The Problem

Industrial ERP systems (like SAP) suffer from "dirty data." Product descriptions are inconsistent (e.g., "Screw M12" vs "M12 Screw Steel"). Standard search algorithms fail to match these duplicates. We needed a way to normalize this text into a strict Schema, but standard LLMs hallucinated non-existent fields.

Training Performance

Live Metrics

Hover over the chart to see how the model's loss decreased (learning) while the JSON schema accuracy rose to near-perfection.

Inference Results

"Zylinderkopfschraube DIN 912 M 12 x 50 - 8.8 verzinkt"
{ "category": "Screw", "norm": "DIN 912", "diameter": "M12", "length": "50mm", "grade": "8.8", "coating": "Zinc" }

Tech Stack

  • Teuken-7B (German)
  • LoRA (Low-Rank Adaptation)
  • PEFT Library
  • NHR "Alex" GPU Cluster
  • DeepSpeed (Stage 2)
  • Hugging Face Accelerate
JSON Parse Rate 100%
Field Accuracy 93.75%