Skip to content

Worked example · Evaluation

Does fine-tuning a small model actually help?

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 Swiss FINMA circulars with a citation to the exact margin number it rests on. The result: fine-tuning plus retrieval doubles the production bar — 22.4% to 45.4%.

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 circulars.

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 Rz in the correct circular. 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 doubles the production bar.

183 human-verified questions (101 German, 82 English, 10 circulars). Brackets are 95% confidence intervals.

Answer correct Answer + citation (production bar) Full system (D)
A · Base
37.2%
0%
B · + Retrieval
63.4%
22.4%
C · Fine-tuned
44.8%
1.6%
D · Fine-tuned + retrieval the full system
80.9%
45.4%
0%25%50%75%100%
Read down the production-bar row: it is near-zero without retrieval (A, C), reaches 22.4% with retrieval alone (B), and only fine-tuning and retrieval together reach 45.4% (D, in red). Full breakdown below.
Metric A · base B · +RAG C · tuned D · tuned+RAG
Answer + citation (production bar) 0.0% 22.4% 1.6% 45.4%
Answer correct 37.2% 63.4% 44.8% 80.9%
Citation correct 1.1% 39.9% 2.2% 50.3%
Emits the required format 2.2% 0.0% 100% 100%

What each layer buys

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

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

Retrieval enables citation

Citation-correct goes from ~1% to ~40–50% only when retrieval is on. A model cannot cite a margin 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 (63% with RAG) but emits the required, machine-parseable format just 2% of the time — 0% once retrieval fills the context and distracts it. The fine-tune nails the format 100% of the time. That gap is most of B → D.

The two compound

Retrieval alone adds 22.4 points; fine-tuning alone adds 1.6. Together they add 45.4 — nearly double the sum of the parts. That interaction is the finding, and it's why we measure before we build.

It looks like this

Same question, same retrieved text.

The base model isn't stupid — it's unreliable in shape. Reliability of shape is what separates a demo from a system.

Q. For banks, securities firms and financial groups, what is the minimum required equity capital at which the fully qualitative and quantitative requirements apply? (gold: CHF 10 billion, Rz 6, FINMA-RS 2010/01)

Base + RAG
CHF 10 billion

Right answer, no citation → fails

Fine-tuned + RAG
ANSWER: CHF 10 billion
SOURCE: FINMA-RS 2010/01 Rz 6

Passes

The honest limits

What this doesn't claim.

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

45% is a floor, not a ceiling. One small model, 4-bit, a single modest LoRA run on ~500 examples, deliberately plain BM25 retrieval. Every one of those is a lever a real Pilot pulls.

Structured answers are the weak spot. Fine-tuned+RAG scores ~47–50% on numbers, durations and yes/no but 27% on multi-item list questions. Complete-set extraction is the obvious next target.

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 decision rests on measurement, not a promise.

FAQ

About this report.

Is this on your customers' data?
No — deliberately. It runs entirely on public FINMA circulars so anyone can reproduce it, and so we can publish it. In a Pilot the identical method runs on your documents, inside your environment, and the report stays yours.
Why is 45% the headline and not something higher?
Because it's honest. Short-form answer plus exact citation is a hard bar, and the setup is deliberately modest — one 8B model at 4-bit, one LoRA run, plain retrieval. The value of the report is the method and the deltas between arms, not a leaderboard number.
How is this scored — is there an LLM judge?
No LLM judge anywhere. FINMA circulars number every normative paragraph (Randziffer, Rz), so a correct citation is exactly checkable. Answer accuracy is string/number matching against a verified gold answer; citation is an exact Rz match. Every number is recomputed from frozen run records.

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.