File size: 9,467 Bytes
436b8d3 ba89a90 436b8d3 ba89a90 436b8d3 ba89a90 436b8d3 c6ab04e 436b8d3 c6ab04e 436b8d3 c6ab04e 436b8d3 c6ab04e 436b8d3 c6ab04e 436b8d3 c6ab04e 436b8d3 c6ab04e 436b8d3 c6ab04e 436b8d3 c6ab04e 436b8d3 c6ab04e 436b8d3 c6ab04e 436b8d3 c6ab04e 436b8d3 c6ab04e 436b8d3 c6ab04e 436b8d3 c6ab04e 08962b1 c6ab04e 08962b1 436b8d3 c6ab04e 436b8d3 c6ab04e 436b8d3 830eff2 436b8d3 830eff2 436b8d3 830eff2 436b8d3 c6ab04e 436b8d3 c6ab04e 436b8d3 c6ab04e 436b8d3 c6ab04e 436b8d3 c6ab04e 436b8d3 c6ab04e 08962b1 c6ab04e 08962b1 c6ab04e 08962b1 436b8d3 c6ab04e 436b8d3 c6ab04e 08962b1 436b8d3 0f176b0 436b8d3 08962b1 c6ab04e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 | ---
language:
- de
- en
- fr
- es
- ar
- fa
- it
- sv
- ru
- zh
license: apache-2.0
library_name: transformers
pipeline_tag: image-text-to-text
tags:
- german
- deutsch
- ocr
- vision
- document-ai
- invoice
- rechnung
- structured-extraction
- json-extraction
- kie
- ollama
- vllm
- llama-cpp
- apache-2.0
inference: true
datasets:
- neuralabs/german-synth-ocr
- Aoschu/German_invoices_dataset_for_donut
base_model:
- Qwen/Qwen3.5-2B
new_version: Keyven/german-ocr-3
---
<p align="center">
<img src="https://app.german-ocr.de/icon.png" alt="German-OCR-3" width="140" height="140" />
</p>
<h1 align="center">German-OCR-3</h1>
<p align="center"><strong>Deutsche Vision-OCR. Kompakt. Lokal. Open Source.</strong><br/>
<sub>Aus deutschem Dokument-Bild → strikt validiertes JSON. In unter 60 Sekunden lokal lauffähig.</sub></p>
<p align="center">
<a href="https://german-ocr.de"><img alt="Site" src="https://img.shields.io/badge/site-german--ocr.de-3B82F6?style=flat-square&labelColor=0B1220"/></a>
<a href="https://ollama.com/Keyvan/german-ocr-3"><img alt="Ollama" src="https://img.shields.io/badge/Ollama-Keyvan%2Fgerman--ocr--3-F59E0B?style=flat-square&labelColor=0B1220"/></a>
<a href="https://github.com/Keyvanhardani/German-OCR-3-Dev"><img alt="GitHub" src="https://img.shields.io/badge/GitHub-source-181717?style=flat-square&labelColor=0B1220"/></a>
<a href="#license"><img alt="License: Apache 2.0" src="https://img.shields.io/badge/License-Apache_2.0-22C55E?style=flat-square&labelColor=0B1220"/></a>
<img alt="Language" src="https://img.shields.io/badge/lang-Deutsch-3B82F6?style=flat-square&labelColor=0B1220"/>
<img alt="Hallucination" src="https://img.shields.io/badge/Halluzination-0%25-22C55E?style=flat-square&labelColor=0B1220"/>
</p>
---
## ⚡ At a glance
<table align="center">
<tr>
<td align="center" width="180"><h2>100 %</h2><sub>Gültiges JSON</sub></td>
<td align="center" width="180"><h2>95 %</h2><sub>Sender korrekt</sub></td>
<td align="center" width="180"><h2>0 %</h2><sub>Halluzination</sub></td>
<td align="center" width="180"><h2>5.0 s</h2><sub>Latenz / Doc</sub></td>
</tr>
</table>
<p align="center"><sub>Auf <strong>200+ echten anonymisierten deutschen Rechnungen</strong> (Default-Edition, 2.7 GB)</sub></p>
---
## Was ist German-OCR-3?
**German-OCR-3** ist eine kompakte, schnelle und voll lokal lauffähige **Vision-OCR-Distribution für deutsche Geschäftsdokumente** — Rechnungen, Briefe, Formulare, Quittungen, Bescheide. Aus dem Bild kommt **strikt validiertes JSON** nach unserem deutschen Extraktions-Schema. Ohne Cloud-Pflicht, ohne Vendor-Lock-in.
Zwei Editionen, beide Apache 2.0, beide unter 3 GB:
| Edition | Ollama | Größe | Zielhardware | Stärke |
|---|---|---:|---|---|
| **Nano** | `Keyvan/german-ocr-nano` | **1.0 GB** | CPU · Edge · Mobile | „läuft überall" |
| **Default** ⭐ | `Keyvan/german-ocr-3` | **2.7 GB** | 4–6 GB VRAM | beste Field-Erkennung |
> **Fine-tuned adapter** für deutsche Geschäftsdokument-Extraktion. Apache 2.0.
---
## 📊 Praxistest — 200+ echte deutsche Rechnungen (anonymisiert)
<p align="center">
<img src="https://huggingface.co/Keyven/german-ocr-3/resolve/main/charts/02_ionos_validity.png" alt="Praxistest" width="820"/>
</p>
| Edition | Valid JSON | Sender korrekt | **Halluzination** | Latenz |
|---|---:|---:|---:|---:|
| `Keyvan/german-ocr-nano` | 84 % | 79 % | **0 %** | 6.6 s |
| **`Keyvan/german-ocr-3`** ⭐ | **100 %** | **95 %** | **0 %** | **5.0 s** |
**Keine "Mustermann"-Defaults.** German-OCR-3 liest echte Firma, Kundenadresse, Produkte, Beträge — statt zu raten.
---
## 📐 Größenvergleich
<p align="center">
<img src="https://huggingface.co/Keyven/german-ocr-3/resolve/main/charts/01_size_vs_competitors.png" alt="Modellgrößen" width="820"/>
</p>
`german-ocr-3` (2.7 GB) ist **6× kleiner** als ein typischer 7B-OCR-VLM. Läuft auf einer **8 GB-Gaming-GPU** oder über CPU auf einem normalen Laptop.
<p align="center">
<img src="https://huggingface.co/Keyven/german-ocr-3/resolve/main/charts/04_latency.png" alt="Latenz" width="620"/>
</p>
---
## 🚀 Quickstart
### Ollama (empfohlen, eine Zeile)
```bash
ollama pull Keyvan/german-ocr-3
ollama run Keyvan/german-ocr-3 "Extrahiere die Rechnung im Bild als JSON." ./meine_rechnung.png
```
<details>
<summary><b>Beispiel-Output (anonymisiert, aus Praxistest)</b> — klicken zum Aufklappen</summary>
```json
{
"document_type": "invoice",
"language": "de",
"invoice_number": "100137xXXXXX",
"invoice_date": "2024-01-22",
"due_date": "2024-01-27",
"sender": {
"name": "IONOS SE",
"address": "Elgendorfer Str. 57, 56410 Montabaur",
"vat_id": "DE81556XXX",
"iban": null
},
"recipient": {
"name": "Firma e.K.",
"address": "Muster Straße 32, 80335 München",
"customer_id": "5835XXX"
},
"line_items": [
{"position": 1, "description": "Mail Business 1 Liz.", "quantity": 1,
"unit": "Monat", "unit_price_net": 4.20, "amount_net": 4.20, "vat_rate": 19}
],
"amount_net": 4.20,
"amount_vat": 0.80,
"amount_total": 5.00,
"currency": "EUR",
"notes": ["Entsprechend Ihrem SEPA-Lastschriftmandat ..."]
}
```
</details>
### Python (via Ollama HTTP API)
```python
import base64, json, requests
from pathlib import Path
b64 = base64.b64encode(Path("rechnung.png").read_bytes()).decode()
resp = requests.post("http://localhost:11434/api/generate", json={
"model": "Keyvan/german-ocr-3",
"prompt": "Extrahiere die Rechnung im Bild als JSON.",
"images": [b64],
"stream": False,
"options": {"temperature": 0, "num_ctx": 32768},
})
data = json.loads(resp.json()["response"])
print(json.dumps(data, indent=2, ensure_ascii=False))
```
### Bundle herunterladen
```bash
huggingface-cli download Keyven/german-ocr-3 --local-dir ./german-ocr-3
# Enthält: Modelfile · JSON-Schemas · System-Prompt · GGUF-Quants · Charts
```
### llama.cpp (GGUF direkt)
```bash
llama-cli -m ./german-ocr-3/german-ocr-3-Q4_K_M.gguf \
--system-prompt-file ./german-ocr-3/system_prompt.txt \
-p "Extrahiere die Rechnung als JSON:" --temp 0
```
---
## 📚 Trainings- und Evaluations-Datensätze
| Datensatz | Umfang | Typ |
|---|---|---|
| [`neuralabs/german-synth-ocr`](https://huggingface.co/datasets/neuralabs/german-synth-ocr) | 4 500+ | Deutsche OCR-Samples (synthetisch, Apache-2.0) |
| [`Aoschu/German_invoices_dataset_for_donut`](https://huggingface.co/datasets/Aoschu/German_invoices_dataset_for_donut) | 129 | Echte deutsche Rechnungen (Donut-Format) |
| Eigenes synthetisches DE-Rechnungs-Set | 100 | Rechnungen mit Golden-JSON, deterministisch generiert |
| Anonymisierter DACH-Praxistest | 200+ | Echte Rechnungen verschiedener DACH-Anbieter (intern, DSGVO) |
---
## 🎯 Zielgruppen
- **Solo-Builder & Indies** — deutsche Dokumente lokal extrahieren, ohne Cloud-OCR-Kosten.
- **DACH-KMU mit Datenschutz-Anspruch** — lokal / on-prem hosten.
- **Agenturen & Studios** — Open-Source-Fundament unter der eigenen Pipeline.
Wer es **gemanagt** und mit noch größeren Modellen will:
> 🌐 **[german-ocr.de](https://german-ocr.de)** — gehostete deutsche OCR-API mit Premium-Modellen, höherer Genauigkeit, ohne eigene Hardware. Daten bleiben in der EU.
---
## ⚠️ Limitations
- Optimiert für **deutsche** Dokumente — andere Sprachen keine Garantie.
- Beste Qualität bei klaren, hochauflösenden Scans/Fotos.
- Handschriftliche Dokumente: nur begrenzt.
- Bei kritischen Vorgängen (Buchhaltung, Recht): **immer Human-in-the-Loop**.
---
## 🙏 Credit & Attribution
German-OCR-3 baut auf der hervorragenden Arbeit des **Qwen-Teams bei Alibaba Group** auf. Die zugrundeliegende Vision-Language-Architektur stammt aus der **Qwen 3.5 Small Series**, veröffentlicht unter Apache License 2.0. Ohne die offene Forschung und die saubere Veröffentlichung der Qwen-Weights wäre dieses Projekt nicht möglich.
- **Qwen 3.5** — https://huggingface.co/Qwen · https://qwen.ai
- **Apache License 2.0** (Weights) — © 2025–2026 Qwen Team, Alibaba Group
- **Qwen2.5-VL Technical Report** — [arXiv:2502.13923](https://arxiv.org/abs/2502.13923)
Vollständiger Attribution-Text in [`NOTICE`](NOTICE).
---
## <a id="license"></a>📄 License
**Apache License 2.0** für die gesamte German-OCR-3-Distribution (Modelfiles, System-Prompt, Schemas, Docs, GGUFs).
---
## 📑 Citation
Wenn du German-OCR-3 in Forschung oder Produktion verwendest, zitiere bitte **beides** — unsere Distribution und die Qwen-Basisarbeit:
```bibtex
@misc{german_ocr_3_2026,
title = {German-OCR-3: A compact German document-OCR distribution},
author = {Hardani, Keyvan},
year = {2026},
url = {https://github.com/Keyvanhardani/German-OCR}
}
@misc{qwen35_2026,
title = {Qwen 3.5 Small Series},
author = {{Qwen Team, Alibaba Group}},
year = {2026},
howpublished = {\url{https://huggingface.co/Qwen}},
note = {Apache License 2.0}
}
@article{qwen25vl_2025,
title = {Qwen2.5-VL Technical Report},
author = {{Qwen Team, Alibaba Group}},
journal = {arXiv preprint arXiv:2502.13923},
year = {2025}
}
```
---
## 👤 Author
**Keyvan Hardani**
· Website: [keyvan.ai](https://keyvan.ai)
· LinkedIn: [linkedin.com/in/keyvanhardani](https://linkedin.com/in/keyvanhardani)
· GitHub: [@Keyvanhardani](https://github.com/Keyvanhardani)
· Hosted Premium: [german-ocr.de](https://german-ocr.de)
|