LLM adaptation and optimisation
Fine-tuning and quantisation.
Your model,
better adapted.
A more consistent response format. A task handled better. A model that needs less memory.
These goals call for different methods. Pepro helps you choose, experiment and compare results on your use cases, up to preparing deployment.

Useful performance is measured
on your task.
AI engineering · Companies, laboratories & R&D teams
From Paris · Delivered in France and abroad
First, the right diagnosis
What would you like
improve in practice?
Before training or converting a model, we look at what is causing it trouble. A better prompt, document retrieval or a different model may already meet the need.
“It can’t find the information in our files.”
RAG. Provide document passages at the moment of the question, with sources you can consult. Explore the document assistant →
“It knows the instruction but doesn’t follow our format well enough.”
Prompt, structured output, then fine-tuning if needed. Check the simplest solutions before training on examples.
“It takes up too much room on our machine.”
Quantisation and deployment settings. Assess weight memory, context and concurrent use, then test the resulting quality.
Fine-tuning:
learning from
your examples.
You show the model what you expect: an input, a desired answer, a way of handling the task. Training adjusts its behaviour from these examples.
- Classify requests into your categories.
- Extract fields in an expected format.
- Follow instructions specific to a repetitive task.
LoRA and QLoRA, simply
LoRA trains small adaptation modules instead of changing every parameter. QLoRA combines this approach with a quantised base model to reduce training memory. Their relevance depends on the model, tools and task.
Quantisation:
representing the weights
with fewer bits.
Weights are the numerical values the model has learned. Storing them at reduced precision can lower the memory needed, without changing the number of parameters.
- Assess deployment on given hardware.
- Reduce the footprint of the model’s weights.
- Compare memory and quality trade-offs.
Lighter does not always mean faster
The result depends on the format, inference engine and hardware. We measure quality, memory and speed under the chosen conditions, without promising the same gain for every model.
A reference point to understand
Fewer bits.
Less space
for the weights.
Choose a size to visualise the theoretical effect of precision. The calculation covers the weights only: it does not size total runtime memory or training memory.
Rule of thumb: parameters × bits ÷ 8.
Decimal GB: 1 GB = 1 billion bytes.
Example: 8 billion parameters
The real requirement is higher. Quantisation metadata, layers kept at another precision, context memory (KV cache) and the inference engine add requirements. Training needs other resources.
Size a configuration with Pepro →The work starts before the compute
Good examples,
beyond simply adding more
of data.
A model can learn your inconsistencies as easily as your good practices. Data preparation is part of Pepro’s service.
Define the expected answer
We clarify the task, formats and edge cases with your business experts. Examples must reflect real work.
Prepare usable data
Usage rights, duplicates, sensitive data, annotations and contradictions are reviewed before training.
Keep cases for evaluation
Examples kept separate from training are used to check whether the model generalises. Duplicates between sets must be checked.
Also compare what degrades
Progress on one task can hide regressions elsewhere. We keep tests for the behaviours to preserve.
A decision based on tests
Before / after.
Under the same conditions.
Our teams work in particular on quantising large models. The goal: reduce hardware requirements while preserving useful performance as much as possible. Checking this requires a clear comparison.
Is the task done better?
Business criteria, format compliance, difficult cases and error review with your team.
What is actually consumed?
Observed memory, hardware, precision and runtime parameters recorded.
How does the service behave?
Response time, throughput and stability under the agreed context and load.
An average alone is not enough: we also document observed errors, versions used and the limits of the comparison.

From the model to the environment that runs it, the choices must work together.
What Pepro takes care of
An experiment.
And a usable next step.
We connect adaptation, resources and deployment. You can call on us for a targeted study or broader support, depending on your starting point.
A scope and a comparison baseline
Model, licence, task, datasets, compute resources and acceptance criteria are identified with you.
Deliverables defined before starting
Depending on the assignment: adapter or converted model, configuration, evaluation results, documentation and handover instructions.
Moving to your environment
We can consider preparing the GPU infrastructure, deployment and technical training for your team.
Budget and schedule
A first test
to inform what comes next.
The volume and condition of the data, the model, the tests needed and the deployment environment determine the cost. The quote separates data preparation, experimentation, compute, integration and any follow-up.
We agree on the scope of the first test and how to read its results. Its outcome may lead to continuing, changing method or staying with the initial solution.
Define my first test with Pepro ↗The starting questions
Before launching
a training run.
The method comes after the need. Here are the points we clarify to avoid starting from a wrong assumption.
Our guide to RAG, fine-tuning and quantisation (in French) →How many examples are needed for fine-tuning?
There is no universal number. Task complexity, variety of cases, annotation quality and the starting model all matter. We review your examples and evaluation criteria to define a first scope, then judge whether it is worth preparing more.
Can we just train the model on all our PDFs?
A collection of PDFs is not, on its own, a set of examples for supervised training. You need to define what the model should learn and prepare the data accordingly. To find information that can be updated, with its sources, a RAG document assistant may be a more direct route.
Does quantisation reduce quality?
It changes the precision of representations and can degrade some behaviours. The effect depends on the model, method and task. We compare answers on representative cases; neither identical quality nor a speed gain is taken for granted.
Can a 4-bit model be adapted with QLoRA?
QLoRA makes it possible to train adapters with a quantised base model, in a compatible environment. This does not mean any 4-bit file can be trained directly. Model, format, libraries and hardware must be reviewed together.
Does fine-tuning eliminate hallucinations?
No. It can improve a task or instruction following, without guaranteeing that every answer is correct. Difficult cases, unsupported answers and behaviours to preserve are among the points to evaluate.
Do we need our own GPU?
Not necessarily. Compute can be organised on existing hardware, infrastructure to prepare or hosted resources chosen for the project. We review access, costs, data and deployment constraints before proposing a solution.
What does my team receive at the end?
Deliverables are specified in the quote: depending on the assignment, adapter or converted model, parameters, test results, documentation and deployment preparation. The model licence and data rights are reviewed. Onboarding or training can complement the handover.
Let’s start from your problem
What would you like your model
do better?
Describe the task, the model you use if you have one and the constraint holding you back. Pepro helps you define what to test first.
Or let’s talk on 09 88 06 54 06
Illustrative photographs: Jakub Pabis and cottonbro studio / Pexels. They do not show a Pepro installation.
