AI Model Training · CPT · SFT · Datasets · Evaluation

Your data.
Your domain.
Your model.

We fine-tune open-weight models on your company's data — manuals, tickets, contracts, processes, terminology — and prove the gain on a frozen benchmark before anything ships. Trained in the EU. You get the weights.

20+ open models on Hugging Face Trained in the EU You own the weights
trainer · qwen3.6-35b-a3b · slovenian
$ bash scripts/run_cpt.sh # QLoRA r256 step 2405/2405 loss 1.42 → 1.34 ✓ adapter saved $ bash scripts/run_sft_autoresume.sh # chat + MT · 2 ep $ python eval/eval_llm_4bit.py --model outputs/merged arc_easy 0.71 → 0.78 boolq 0.83 → 0.89 avg (7) 0.623 → 0.654 +3.1 pts · 6/7 tasks won $ bash scripts/merge_lora.sh && quantize Q4_K_M ✓ GGUF ready
held-out perplexity · CPT 6.996 → 3.438
6.996 3.438 0 1.78 B tok
base vs tuned Slovenian-LLM-Eval
arc_easy0.710.78+7
boolq0.830.89+6
hellaswag0.590.62+3
sl→en BLEU30.935.0+4.1
ASR WER46.822.3−52 %
20+Open models published on Hugging Face, each with a card, provenance and numbers
1.78BTokens in our largest continued pre-training run (full-parameter, 4B model)
−52 %Speech-recognition word error after fine-tuning on 962 h of audio
140 kDomain training records engineered, QC-gated and licence-tracked
The gap

General models have never seen your business.

Frontier APIs are excellent at the average case. Your tickets, manuals, contracts and processes are not the average case — and the last 20 % of quality is where the value is.

01 — KNOWLEDGE

It has not read your documents

Your product manuals, support history, procedures and terminology exist only inside your company. No off-the-shelf model has seen them — so it guesses, confidently.

02 — RELIABILITY

Prompts get you to 80 %

Exact output formats, house terminology, tool calls that work every time, answers in your language — the last 20 % is trained in, not prompted in.

03 — SOVEREIGNTY & COST

Per-token APIs don't scale, or comply

Sending customer data to a foreign API is a compliance and lock-in problem. A small model trained on your data runs on your own servers at a flat cost — and the weights are yours.

What matters Public API Prompting + RAG Fine-tuned with us
Data stays in the EU, under your control~
Your terminology and output formats, every time~~
Reliable tool calls into your systems~~
Facts that change dailyretrieval belongs here — we combine both
Flat cost at scale, runs offline~
You own the model
What we do

From raw data to a model you own.

Six capabilities, one pipeline. Take any of them alone or the whole chain end-to-end. Every one has been exercised on a model you can download.

01

Continued pre-training (CPT)

Teach a base model a whole language or a domain corpus — billions of tokens of cleaned, deduplicated, licence-tracked text.

  • Full-parameter training for dense models; high-rank QLoRA for 30B+ mixture-of-experts
  • Perplexity gates on held-out text after every checkpoint
  • Retention mixes so the model keeps what the base did well
  • Packed sequences, crash-resilient multi-day runs
1.78 B tokens → held-out perplexity −51 % · Sokol 4B
02

Supervised fine-tuning (SFT)

Shape behaviour: chat, translation, tool calling, reasoning, identity — in the model's own template, the one your serving stack renders.

  • Native function calling with multi-turn call → result → answer loops
  • Reasoning distillation from a stronger teacher, judge-filtered
  • Identity, tone and policy tuning
  • Sequential re-tunes that build on earlier adapters
+3.1 pts Slovenian-LLM-Eval · 6/7 tasks won · +4.1 BLEU sl→en
03

Dataset engineering

Training-ready corpora with provenance — not scraped piles. Half of every program is here.

  • One schema, quality scoring, exact + embedding dedup, deterministic splits
  • Licence gating: commercial mix and research mix from the same sources
  • Teacher distillation with judge filtering; synthetic dialogues from verified blocks
  • Domain translation with terminology freezing — codes, names, IDs survive intact
  • Exports for CPT, SFT and preference training (DPO/GRPO) + frozen hold-outs
140,063 QC-gated medical records · 35-source cybersecurity registry · 49,772 distilled traces
04

Training runs as a service

Bring data, get back a model. Or bring nothing and we build the data too.

  • GPU capacity we control — own and vetted EU partner clouds — or your cloud / air-gapped hardware
  • Auto-resume wrappers and checkpoint discipline
  • Merged bf16 weights, LoRA adapters and GGUF quantizations delivered together
  • Model card with full provenance and the numbers behind every claim
20 models and 4 datasets published on Hugging Face
05

Evaluation flywheel

"Better" is measured, not asserted. The benchmark is defined before the data is touched.

  • Frozen, versioned benchmark sets — never edited, only extended
  • Objective track (accuracy, BLEU/chrF, WER/CER) + multi-judge track
  • Regression-flagging scoreboard across every model version
  • Go/no-go on numbers before anything ships
every model in Results was scored base-vs-tuned on the same frozen set
06

Deployment & speech

From a 4B model on a laptop to a 35B mixture-of-experts on a workstation — with vision, tools and voice.

  • GGUF Q4–Q8 for llama.cpp / LM Studio; vLLM and transformers recipes
  • Image input and tool use working in production, not just in the demo
  • Streaming speech recognition and text-to-speech fine-tuning
  • Optional hosting or on-site install that runs unplugged
Slovenian streaming ASR — WER 46.8 % → 22.3 %, runs on CPU
How it works

Five steps, one number at the end.

A first measured model lands within weeks, not quarters. Every phase ends with something you can check.

  1. Scope

    Define the job and the eval

    The users, the task, the frozen eval set. An audit of your data and its licences. A written go/no-go criterion.

  2. Data

    Build the corpus

    Clean, translate, distil and mix. Retention slices protect what the base already does well.

  3. Train

    CPT → SFT → re-tunes

    Each checkpoint gated by a quick eval so bad runs die early and good ones get the budget.

  4. Evaluate

    Base vs tuned vs alternatives

    Head-to-head on the frozen benchmark, including the API or RAG setup you were considering.

  5. Ship

    Weights, GGUF, card, recipe

    Hand-over workshop. Hosting or an on-site install if you want it. No dependency on us remains.

Results

Everything below is public.

Every model and dataset sits on Hugging Face with a card, its provenance and its numbers. Slovenian is our reference implementation — the same pipeline runs on your domain.

Qwen3.6-35B-A3B · base vs CPT+SFT (Slovenian)

Slovenian-LLM-Eval · acc_norm · n=100/task · 4-bit eval
TaskBaseTunedΔbase ░ tuned ▓
arc_easy0.710.78+0.07
arc_challenge0.480.51+0.03
hellaswag0.590.62+0.03
piqa0.650.64−0.01
openbookqa0.480.51+0.03
winogrande0.620.63+0.01
boolq0.830.89+0.06
average0.6230.654+0.031
CPTSFTtool-callingvisionGGUF

Qwen3.6-35B-A3B Slovenian

A 35B mixture-of-experts (3B active) taught Slovenian through continued pre-training, then chat and translation SFT, then a tool-calling re-tune that kept the language gains. Vision and native function calls included.

+3.1pts acc_norm
6/7tasks won
+4.1BLEU sl→en
full-parameter CPTSFTreasoning distillationon-device

Sokol 4B — on-device assistant

A dense 4B model pre-trained on 1.78 billion Slovenian tokens, then fine-tuned on ~250k examples of chat, translation, tool calling and teacher-distilled reasoning. One repo, full weights plus Q4–Q8 GGUF. Runs on a laptop.

−51 %perplexity
+0.047acc_norm avg
6/7tasks won
dataset engineeringdomain MTeval suiteresearch

Vran 27B — biomedical research model

We built the corpus first: 140,063 Slovenian medical instruction records, machine-translated from a provenance-tracked English gold set with a purpose-built medical translation model and a four-layer QC gate, plus a dedicated eval suite. Then we trained the model. Research use.

140 kQC-gated records
27 Bparameters
3datasets released
ASRfull fine-tunestreamingCPU deploy

Slovenian streaming speech recognition

Full fine-tune of NVIDIA's Nemotron 0.6B streaming recogniser on 962 hours of Slovenian speech, scaled up in versions with no catastrophic forgetting. Word-error rate halved on studio speech and cut by two thirds on multi-domain audio. Deployed on CPU as GGUF.

22.3 %WER studio (from 46.8)
23.4 %WER multi-domain (from 67.7)
962 htraining audio
TTSLoRAvoice stack

Slovenian text-to-speech voices

LoRA fine-tunes of two open speech generators on Slovenian studio speech, closing the loop of a fully local speech-in → speech-out assistant with camera vision and tool use — served from one command in the browser.

2open TTS models
100 %runs locally
datasetsCPT corpusSFTbenchmark

Published datasets

A cleaned continued-pre-training corpus, a 120,000-pair bidirectional EN↔SL translation SFT set, the medical instruction corpus and its evaluation suite — each with a card that states sources, licences and how it was built.

Use cases

What companies fine-tune with us.

Typical starting points. Most programs combine two or three.

Customer support & service

Assistants trained on your ticket history and manuals: consistent answers, your tone, escalation when unsure.

Documents & extraction

Contracts, invoices, forms, reports: structured extraction, classification and summaries in your schema, in your language.

Internal knowledge

Procedures, standards, ISO documentation, engineering notes — a colleague that has actually read them, offline.

Agents & automation

Reliable tool calling into your ERP, CRM or ticketing API. Trained on your function signatures, tested on pass rate.

Speech & voice

Call transcription with your vocabulary, voice bots and dictation — recognition and synthesis that run on your own servers.

Language adaptation

Slovenian or any under-served language: continued pre-training and instruction tuning so the model is fluent before it is expert.

Ownership

Your data stays in the EU. The model leaves with you.

We train on GPU capacity we control — our own and vetted partner clouds inside the EU. Either way your data is handled under NDA and never leaves European jurisdiction.

What you receive

  • ModelMerged weights + LoRA adapters
  • DeploymentGGUF / vLLM build + serving recipe
  • DataEvery dataset we built, with provenance
  • EvaluationFrozen benchmark + base-vs-tuned scoreboard
  • DocumentationModel card, data card, training configs
  • Hand-overWorkshop; optional hosting or on-site install

The rules

  • Your data is handled under NDA, on infrastructure we control, inside the EU. GDPR is the default, not an add-on.
  • Air-gapped training for the most sensitive data: we run with the network cable unplugged and hand over the drive.
  • Open-weight bases only — Qwen, Gemma, Llama, Mistral, Nemotron — with the licence of every base and every data source tracked.
  • You receive the weights, the adapters, the datasets, the eval set and the provenance record. No dependency on us remains unless you want one.
  • Everything is versioned — data, configs, checkpoints, scores — so a result can be reproduced a year later.
Engagement

Start small, measure, then commit.

Three ways to work with us. Most teams start with the pilot and decide on the numbers.

End-to-end

Full program

CPT → SFT → evaluation → deployment for a production model, including the dataset engineering and the serving recipe.

  • Corpus build, translation or distillation as needed
  • Multi-stage training with checkpoint gates
  • GGUF / vLLM delivery, model card, hand-over workshop
Discuss a program
Monthly cadence

Flywheel retainer

New data in, retrain, re-evaluate, ship — every month, with the scoreboard as the contract.

  • Versioned benchmark that grows with your use cases
  • Regression alerts before users notice
  • Base-model upgrades when a better open model appears
Talk about a retainer

GPU time is also available on its own for teams that bring their own pipeline.

FAQ

Straight answers.

Who owns the model?

You do. Weights, adapters, the datasets we build for you, the eval set and the provenance record are delivered to you. We keep nothing you don't want us to keep.

Fine-tuning or retrieval (RAG)?

Usually both. Retrieval is right for facts that change daily; fine-tuning is right for behaviour — formats, terminology, tool use, tone, and making a small model do what only a large one could. The pilot tests which one actually moves your metric before you commit.

Which base models do you work with?

Any open-weight model whose licence fits your use: Qwen, Gemma, Llama, Mistral for text; Nemotron and Parakeet for speech recognition; open TTS models for speech synthesis. We track the licence of every base and every data source, and we will tell you if a combination is not commercially usable.

Which languages?

Any. Most industry programs run in the customer's language on top of a strong multilingual base; Slovenian is our proven case end to end. For under-served languages the same recipe — continued pre-training, fine-tuning, frozen evaluation — applies, and we assess honestly how much usable text exists before promising anything.

How much data do I need?

For fine-tuning, a few thousand high-quality examples already move the needle. For continued pre-training, hundreds of millions of tokens. What you don't have, we can build, translate or distil — that is half of what we do.

Can the model run offline?

Yes. Our 4B models run on a laptop; 27–35B models on a single workstation GPU. Everything we ship works with the network cable unplugged.

Does this help with the EU AI Act?

With the documentation side, yes: every model ships with a data card, provenance record, training configs and evaluation results — the artefacts transparency and record-keeping obligations ask for. We are not lawyers; we give your compliance team what it needs.

How long does it take?

A measured pilot model within weeks. A full program depends on data readiness; you get a dated schedule after the scoping week, and every phase ends with a number you can check.

Send us fifty examples.

Show us fifty real prompts and answers from your domain. Within a week we'll tell you what a model trained on your data could do — and exactly what it would take.

MediaAtlas d.o.o. · Sevnica, SloveniaNDA by defaultEU jurisdiction