---
title: README
emoji: 🧭
colorFrom: gray
colorTo: yellow
sdk: static
pinned: false
---

# MediaAtlas — LLM fine-tuning on your data, trained in the EU

**Your data. Your domain. Your model.** We fine-tune open-weight language models on a company's own data — manuals, tickets, contracts, processes, terminology — and prove the gain on a frozen benchmark before anything ships. Trained on GPU capacity inside the EU. You get the weights.

**→ [Service overview](https://mediaatlas.si/ai-training.html) · [Pricing](https://mediaatlas.si/ai-training-pricing.html) · [Book a call](https://cal.com/tfius-tfius) · info@mediaatlas.si**

## What we do

- **Continued pre-training (CPT)** — full-parameter or high-rank QLoRA, for domain or language adaptation
- **Supervised fine-tuning (SFT)** — chat, translation, native tool calling, reasoning distillation, identity
- **Dataset engineering** — schema, dedup, quality scoring, licence gating, teacher distillation, domain MT
- **Evaluation flywheel** — frozen benchmark sets, objective + multi-judge tracks, regression scoreboard
- **Deployment & speech** — GGUF (llama.cpp / LM Studio), vLLM, on-device 4B to 35B MoE; ASR and TTS fine-tunes
- **Flywheel retainer** — monthly retrain, re-score, ship; base-model upgrades when a better open model wins on your eval

## Public reference work (Slovenian)

All models below are published under [texdata](https://huggingface.co/texdata) with model cards, provenance and numbers.

| Model | What |
|---|---|
| [Qwen3.6-35B-A3B-Slovenian](https://huggingface.co/texdata/Qwen3.6-35B-A3B-Slovenian-GGUF) | CPT + SFT Slovenian MoE, chat + EN↔SL translation, GGUF |
| [Sokol-4B-SLO](https://huggingface.co/texdata/Sokol-4B-SLO) | On-device Slovenian assistant, 1.78 B-token full-parameter CPT, tool calling, reasoning |
| [Vran-27B-SLO-BioMed-Research](https://huggingface.co/texdata/Vran-27B-SLO-BioMed-Research) | Slovenian biomedical research model (not for clinical use) |
| [Qwen3.6-27B-slo-med-mt](https://huggingface.co/texdata/Qwen3.6-27B-slo-med-mt-GGUF-Q8) | EN→SL medical translation with terminology freezing |
| [nemotron-3.5-asr-streaming-slovenian](https://huggingface.co/texdata/nemotron-3.5-asr-streaming-slovenian) | Streaming Slovenian ASR, WER 46.8 → 22.3 % |
| [Gepard-Slovenian-TTS](https://huggingface.co/texdata/Gepard-Slovenian-TTS) · [Maya1-Slovenian-TTS](https://huggingface.co/texdata/Maya1-Slovenian-TTS) | Slovenian text-to-speech voices |
| [Qwen3.6-35B-A3B-Tool-Slo](https://huggingface.co/texdata/Qwen3.6-35B-A3B-Tool-Slo-GGUF) | Native tool / function calling in Slovenian |

Datasets: [med-slo-sft](https://huggingface.co/datasets/texdata/med-slo-sft) · [med-slo-eval](https://huggingface.co/datasets/texdata/med-slo-eval) · [slo-pretrain](https://huggingface.co/datasets/texdata/slo-pretrain) · [opus100-en-sl-translate-sft](https://huggingface.co/datasets/texdata/opus100-en-sl-translate-sft)

Language adaptation is the public reference implementation; the same pipeline runs on your support tickets, documents, internal knowledge or agent workflows, in your language.

## How an engagement works

1. **Pilot** — fixed scope, fixed price: one eval set, one dataset, one fine-tune, one report.
2. **Full program** — CPT → SFT → evaluation → deployment, quoted after a scoping week.
3. **Flywheel retainer** — the model stays current as base models and your data move.

Data handled under NDA, inside the EU. Air-gapped training available. Open-weight bases only, every licence tracked.

*MediaAtlas d.o.o. · Kajuhova 11, 8290 Sevnica, Slovenia · [mediaatlas.si](https://mediaatlas.si)*