Work Order Text
Enter a maintenance work order description. The model classifies the failure category and finds the 3 most similar past cases.
Classification
Similar Past Cases
Most semantically similar work orders from the training corpus.
What Makes This Different
Domain Data
Corpus generated with real maintenance vocabulary from 12 years of industrial engineering — HVAC, subsea, manufacturing. Realistic abbreviations, failure taxonomies, and work-order style.
LoRA Fine-Tuning
DistilBERT fine-tuned with LoRA adapters: 1.1% of parameters updated for 94.4% F1 vs. 95.9% full fine-tune, 1.6× faster training. The adapter is 2.8 MB — the full model is 256 MB.
LLM ETL Extraction
GPT-4o-mini with Pydantic structured output lifts failure-mode extraction from 13% to 70% on noisy technician text. The measured lesson from the hybrid extractor: regex self-confidence is a poor escalation signal — production needs a learned gate.