Applied Weights
RESULTS / FROM THE LAB

Real tasks.
Recorded results.

FIRST EXPERIMENT
EXPERIMENT 001 / STRUCTURED EXTRACTION

From claim narratives
to structured data.

We fine-tuned Qwen2.5-7B on 1,000 synthetic insurance claims to turn written narratives into structured records: policy details, incident dates, insured objects, and damage estimates. We then compared its extracted fields with reference answers and a GPT-4o baseline.

Qwen2.5 · 7BQLoRA1,000 training examples
EXTRACTION ILLUSTRATIONSAMPLE / 01
↓UNSTRUCTURED → STRUCTURED
Illustrative examples · prewritten outputs · no live model
FINE-TUNED QWEN · FIELD F186.31%
GPT-4o BASELINE · FIELD F191.91%
Field F1 measures how closely the extracted facts match the reference answers. The fine-tuned open model scored 86.31%, compared with 91.91% for GPT-4o—a gap of 5.60 percentage points.
← Back to Applied WeightsContact us