A LoRA adapter for Mistral-7B-Instruct-v0.2 that converts informal,
messy stakeholder requests into structured engineering tickets
(Title + Description format), specialized for insurance industry
workflows — claims, underwriting, billing, compliance, and policy admin.
Training details
Base model: mistralai/Mistral-7B-Instruct-v0.2
Method: LoRA (r=16, alpha=32, target modules: q/k/v/o_proj) via PEFT,
4-bit NF4 quantized (QLoRA-style) for training on a free Colab T4 GPU
Trainable parameters: 13.6M / 7.26B total (0.19%)
Training data: 15 hand-crafted instruction/response pairs covering
real insurance domain scenarios
6 optimizer steps, 3 epochs — training loss dropped from 3.60 to 1.92
Before / after example
Input: "the system doesnt stop someone from binding a policy in a
state we're not even licensed to sell in"
Base Mistral: Produces a rambling response with inconsistent
structure ("Steps to Reproduce," "Expected Result," closing with
"We appreciate your prompt response on this critical security concern.")
Fine-tuned:
Title: Unauthorized Policy Binding in Unlicensed State
Description: The current system configuration allows users to
bind policies in states where our organization is not authorized...
What changed vs. what didn't
Fine-tuning did NOT make the model "smarter" — it changed its behavior:
consistent Title/Description structure, terse professional tone, no
conversational filler. This is the real distinction between fine-tuning
(behavior/style) and RAG (knowledge) — this project isolates and
demonstrates that difference directly.
Limitations
Trained on only 15 examples / 6 optimizer steps as a learning exercise
Consistently produces Title + Description; the third intended
section (Acceptance Criteria) did not reliably appear even with
extended generation length — likely needs more training examples
or steps to fully lock in the 3-part format
A production version would use hundreds of real historical tickets
