Skip to content

Worked example · Insurance

Does fine-tuning help on Swiss insurance-contract law?

Not an opinion — a measurement. We took Apertus-8B and ran the same four-way ablation we run at the start of every Pilot, on a task where the answer can be scored exactly: answering questions about the Swiss Insurance Contract Act with a citation to the exact article it rests on — with a citation to the exact article it rests on. The result: fine-tuning plus retrieval takes the production bar from 18.8% to 37.5%.

The setup

Four systems, one open model.

Same Apertus-8B, quantized identically to 4-bit, served identically. The only variable is what we added.

A

Base

The open model, as shipped.

B

+ Retrieval

Base model with BM25 retrieval over the law.

C

Fine-tuned

LoRA-tuned on the task, no retrieval.

D

Fine-tuned + retrieval

The full system.

Two things are scored, both deterministically — no LLM judge: is the answer correct, and did it cite the correct article. The production bar is both at once — a right answer with a wrong or missing citation is a failure, because in regulated work it is one.

The result

Fine-tuning plus retrieval is the winning system.

16 verified questions (10 German, 6 French). Brackets in the report are 95% confidence intervals.

Answer correct Answer + citation (production bar) Full system (D)
A · Base
43.8%
6.2%
B · + Retrieval
43.8%
18.8%
C · Fine-tuned
43.8%
0%
D · Fine-tuned + retrieval the full system
75%
37.5%
0%25%50%75%100%
Read the production-bar row: near-zero without retrieval (A, C), 18.8% with retrieval alone (B), and only fine-tuning and retrieval together reach 37.5% (D, in red).
Metric A · base B · +RAG C · tuned D · tuned+RAG
Answer + citation (production bar) 6.2% 18.8% 0.0% 37.5%
Answer correct 43.8% 43.8% 43.8% 75.0%
Citation correct 6.2% 62.5% 6.2% 43.8%
Emits the required format 0% 0% 94% 69%

What each layer buys

The ablation stops anyone — us included — from hand-waving.

Each card pins one mechanism to a number in the table above.

Retrieval enables citation

Citation-correct climbs from 6.2% to 62.5% only when retrieval is on. A small model cannot cite an article number it has never seen — so both no-retrieval arms sit near zero on the production bar, however well-trained.

Fine-tuning enables usable output

The base model often knows the answer (43.8% with RAG) but emits the required, machine-parseable two-line format just 0% of the time. The fine-tune makes it reliable (69%), and lifts citation under retrieval to 43.8%.

Fine-tuning alone is not enough

Arm C — fine-tuned, no retrieval — reaches 94% format but only 6.2% citation and 43.8% answer: the model learns how to answer, not the exact facts. The winning system is the combination, arm D at 37.5%.

The honest limits

What this doesn't claim.

An evaluation you can trust names where it's weak.

37.5% is a floor, not a ceiling. One small model, 4-bit, a single modest LoRA run, deliberately plain BM25 retrieval, a compact verified eval set (n=16). Every one of those is a lever a real Pilot pulls.

Fine-tuning alone can regress. Arm C's answer accuracy sits below the base model — proof that fine-tuning without retrieval is the wrong tool for exact-citation tasks. We publish that rather than hide it.

The method is the product, not the number. Your task, your documents, this exact ablation — run before you commit to production, so the go/no-go rests on measurement, not a promise.

FAQ

About this report.

Is this on our data?
No — deliberately. It runs entirely on the public, pinned Fedlex text of the VVG so anyone can reproduce it. In a Pilot the identical method runs on your policies and correspondence, inside your environment, and the report stays yours.
Why is the base model so weak here?
Insurance-contract questions are nuanced and the base model's parametric answers are unreliable (answer accuracy stays low even with retrieval). That is exactly the case where the fine-tune's format discipline and citation gains matter most — and why we measure before we build.
Is there an LLM judge?
No LLM judge anywhere. The VVG numbers its articles; a correct citation is exactly checkable. Answer accuracy is string/number matching against a verified gold answer; citation is an exact article match. Every number is recomputed from frozen run records with 95% Wilson intervals.

Want this on your task?

Bring one workflow and the documents behind it. We'll build the evaluation set and run this ablation — on your data, in your jurisdiction.