Fine-tuning & Evaluation | Avanzar Solution — Generative AI
Fine-tuning & Evaluation

Teach the model your domain — then prove it got better.

We fine-tune open and hosted models on your own curated data, and we hold every version to a scored evaluation suite. No launch happens because a demo looked good; it happens because the numbers moved on a test set you agreed to.

What It Is

Fine-tuning changes the model. Evaluation tells you if it helped.

Fine-tuning continues a pre-trained model's training on examples of your own work, so it absorbs your terminology, format, and house style rather than being reminded of them in every prompt. Evaluation is the other half of the job: a fixed set of test cases, scored the same way every run, that shows whether version four is genuinely better than version three or just different.

Data curation

Collect, clean, and label real examples

Method choice

Full SFT, LoRA, or preference tuning

Training

Runs tracked, checkpoints kept

Evaluation

Scored on a held-out golden set

Deployment

Shipped behind a versioned endpoint

Monitoring

Live scoring and drift alerts
Evaluation comes before the first training run, not after — otherwise there's nothing to compare the result against.
Before You Tune

Fine-tuning isn't always the right answer

It's the most expensive of the three levers, so we check the cheaper two first — and say so if they're enough.

Try prompting first

If the model can already do the task and just needs clearer instructions or a few examples, prompt engineering gets you there in days, not weeks — and costs nothing to change later.

Then try retrieval

If the gap is missing knowledge — your policies, product data, recent records — RAG fixes it without touching model weights, and stays current when the source documents change.

Fine-tune for behaviour

When the gap is how the model behaves — output format, tone, domain reasoning, or consistency across thousands of calls — that lives in the weights, and tuning is the tool that reaches it.

Tuning Approaches

Four methods, matched to the problem

Which one we reach for depends on how much data you have, how much you can spend on compute, and whether the target is knowledge, format, or judgement.

A curated dataset of prompt and response pairs passing a quality check

Supervised Fine-tuning (SFT)

The workhorse. We assemble prompt-and-ideal-response pairs from your real work, clean out the contradictions, and train the model to reproduce that pattern. Quality of examples matters far more than quantity — a few thousand good pairs beat fifty thousand noisy ones.

Instruction tuning Output formatting Domain language
A frozen base model with small trainable adapter modules attached

LoRA & Parameter-Efficient Tuning

Instead of retraining billions of weights, we freeze the base model and train small adapter layers alongside it. Training costs a fraction of full fine-tuning, adapters are a few megabytes, and you can keep several — one per task or per client — and swap them at request time.

LoRA / QLoRA Multi-adapter serving Low GPU cost
Two candidate answers to one prompt, with the preferred one selected

Preference Tuning (DPO / RLHF)

Some qualities are easier to judge than to write down. Here your reviewers pick the better of two answers, and the model learns from those choices — which is how you shape tone, refusal behaviour, and the sense of what a "good" answer looks like in your context.

Tone & style Safety behaviour Expert feedback loops
A large teacher model distilled into a smaller, faster student model

Distillation & Compression

Once a large model performs well on your task, we use its outputs to train a much smaller one. The student keeps most of the quality at a fraction of the latency and cost — which is what makes high-volume or on-device deployment affordable.

Smaller models Lower latency Quantization

"It feels better" is not a result

Most fine-tuning projects go wrong in the same place: the team tests a handful of prompts by hand, agrees the output looks nicer, and ships. Two months later nobody can say whether the model improved, and the next change is a coin flip.

We build the evaluation set before the first run, keep it frozen, and report every version against it. When a change makes something worse, you find out in the report — not from a customer.

Discuss Your Use Case →
Evaluation

How we measure whether a model is actually better

Every tuned version is scored the same way, on the same cases, and the results sit in one place your team can open.

An evaluation dashboard comparing base and tuned model scores across versions, with a passing test suite and judge scorecard

Golden datasets

A frozen set of real cases with reviewed correct answers, built with your subject-matter experts. It's the yardstick every version is held to.

Task metrics

Exact-match, F1, or structured-output validity where the answer is checkable, so a large part of scoring runs automatically on every commit.

LLM-as-judge

For open-ended answers, a scoring model rates each response against a written rubric — calibrated against human ratings first, so the scores mean something.

Retrieval evals

Where a RAG layer is involved, we score faithfulness, context relevance, and answer groundedness separately — so you know whether a failure was retrieval or generation.

Safety & red-teaming

Adversarial prompts, prompt-injection attempts, and PII leakage checks run as part of the suite — because tuning can quietly weaken a base model's guardrails.

Regression suites

The cases you've already fixed get pinned as tests. Any future version that breaks one fails the run, so improvements don't quietly undo each other.

Our Stack

The stack behind our tuning and eval work

Pick a category to see what we build with and why it's in the stack.

Transformers

Hugging Face Training

PEFT / LoRA

Adapter Training

TRL

DPO & RLHF Training

PyTorch

Core Training Runtime

DeepSpeed / Ray

Distributed Runs
How We Work

From baseline to a version you can trust

Five stages, each ending in a number you can compare against the one before it.

01

Baseline & eval set

We score the untuned model first and build the golden set with your experts.

02

Data curation

Assemble, de-duplicate, and review the training pairs — the step that decides the outcome.

03

Training runs

Several configurations, tracked and versioned, compared on the same held-out set.

04

Evaluation report

Scores, failure examples, safety checks, and a plain recommendation on whether to ship.

05

Ship & monitor

Versioned rollout with live scoring, so regressions surface in the dashboard, not in support tickets.

Not sure whether your model needs tuning or better prompts?

Send us the task and a handful of outputs you're unhappy with. We'll tell you which of the three levers will fix it, what data you'd need, and what the evaluation would look like — before any training starts.

Start a Project →
FAQ

Questions teams ask us before starting

Don't see your question here? Send it over and we'll answer it directly.

Fewer examples than most teams expect, but they have to be good ones. A focused format or tone change often lands with 500 to 2,000 reviewed pairs; broader domain reasoning wants several thousand. Contradictory examples hurt more than missing ones, so curation time usually outweighs collection time.

If the gap is missing knowledge, RAG fixes it and stays current when your documents change. If the gap is behaviour — output format, tone, domain reasoning, consistency across thousands of calls — that lives in the weights and fine-tuning is the right tool. Plenty of systems end up using both.

It can, and that's called catastrophic forgetting. We guard against it by mixing in general examples, favouring adapter-based methods that leave the base weights untouched, and keeping general-capability checks in the evaluation suite so a regression shows up in the report.

A frozen golden dataset built with your experts, automatic metrics where answers are checkable, an LLM judge scored against a written rubric for open-ended cases, safety and prompt-injection checks, and a regression set of every bug you've already fixed.

Only after it's been calibrated. We score a sample by hand first and compare it against the judge's ratings; if they don't line up, the rubric gets rewritten before the judge is used for anything that matters. It's a cost-saving tool, not a replacement for human review.