---
language:
  - de
  - en
license: mit
base_model: deepset/gbert-base
pipeline_tag: text-classification
library_name: transformers
tags:
  - klarki
  - eu-ai-act
  - compliance
  - german
  - text-classification
  - bert
model-index:
  - name: klarki-bert-classifier
    results:
      - task:
          type: text-classification
          name: Text Classification
        dataset:
          name: KlarKI EU AI Act Regulatory Training Data
          type: custom
        metrics:
          - type: f1
            value: 0.954
            name: Macro F1
            verified: false
---

# KlarKI — EU AI Act Article Domain Classifier

> 8-class text classification — maps document chunks to EU AI Act article domains (Articles 9–15 + unrelated)

> [!NOTE]
> Part of **[KlarKI](https://github.com/s4nkar/klarki)** — a local-first EU AI Act + GDPR compliance auditor for German SMEs.
> All inference runs on-device. No data leaves your machine.

---

## Model Overview

| Property | Value |
|---|---|
| Base model | [deepset/gbert-base](https://huggingface.co/deepset/gbert-base) |
| Architecture | Transformers — `BertForSequenceClassification` |
| Parameters | ~110M parameters |
| Languages | German (primary), English |
| Training samples | 5536 train / 981 validation |
| License | MIT |
| Part of | [KlarKI](https://github.com/s4nkar/klarki) audit pipeline |

---

## Quickstart

### Option A — Via KlarKI (recommended)

> [!TIP]
> Use this if you want the full audit pipeline. The download script places all 5 models
> exactly where KlarKI expects them.

```bash
git clone https://github.com/s4nkar/KlarKI-EU-AI-Act-compliance-auditor.git
cd KlarKI-EU-AI-Act-compliance-auditor
pip install huggingface-hub>=0.26.0
python scripts/download_pretrained.py --model bert
./run.sh up
```

### Option B — Direct usage

```python
from transformers import pipeline

classifier = pipeline("text-classification", model="s4nkar/klarki-bert-classifier")
result = classifier("The system must maintain a risk management system throughout the entire lifecycle of the AI system.")
# Output: [{'label': 'risk_management', 'score': 0.97}]
```

---

## Labels

| Label | Description |
|---|---|
| `risk_management` | Article 9 — Risk Management System |
| `data_governance` | Article 10 — Data and Data Governance |
| `technical_documentation` | Article 11 — Technical Documentation |
| `record_keeping` | Article 12 — Record-Keeping |
| `transparency` | Article 13 — Transparency and Provision of Information |
| `human_oversight` | Article 14 — Human Oversight |
| `security` | Article 15 — Accuracy, Robustness and Cybersecurity |
| `unrelated` | Not related to EU AI Act Articles 9–15 |

---

## Evaluation Results

**Overall**

| Macro F1 | Val samples |
|---|---|
| 0.9540 | 981 |

**Per-Class**

| Class | Precision | Recall | F1 | Support |
|---|---|---|---|---|
| `risk_management` | 0.9435 | 0.9512 | 0.9474 | 123 |
| `data_governance` | 0.9593 | 0.9672 | 0.9633 | 122 |
| `technical_documentation` | 0.9680 | 0.9680 | 0.9680 | 125 |
| `record_keeping` | 0.9583 | 0.9426 | 0.9504 | 122 |
| `transparency` | 0.9569 | 0.8952 | 0.9250 | 124 |
| `human_oversight` | 0.9365 | 0.9672 | 0.9516 | 122 |
| `security` | 0.9516 | 0.9593 | 0.9555 | 123 |
| `unrelated` | 0.9593 | 0.9833 | 0.9712 | 120 |

---

## Training Details

| Property | Value |
|---|---|
| Base model | `deepset/gbert-base` |
| Training epochs | 5 (AdamW, early stopping) |
| Batch size | 16 |
| Data split | 85% train / 15% validation, stratified, seed=42 |
| Data generation | Async Ollama-grounded synthesis (phi3:mini) + real regulatory text |
| Optimiser | AdamW |
| Training framework | Docker container (Python 3.11, isolated from host) |

---

## Intended Use

Routing document chunks to the correct article gap analyser inside the KlarKI audit pipeline. Each 512-character chunk is assigned to one of seven article domains or marked `unrelated`.

> [!WARNING]
> This model is a **decision-support tool**, not a substitute for qualified legal advice.
> EU AI Act compliance determinations should always be reviewed by a legal professional.

---

## Limitations

- Trained primarily on German regulatory text; performance may degrade on highly informal language.
- `unrelated` is a catch-all class; very short or ambiguous chunks may be misclassified.
- Designed for 512-character chunks, not full documents.

---

## Citation

```bibtex
@software{klarki2026,
  author    = {Sankar},
  title     = {KlarKI: Local-First EU AI Act and GDPR Compliance Auditor},
  year      = {2026},
  url       = {https://github.com/s4nkar/KlarKI-EU-AI-Act-compliance-auditor},
  note      = {Open-source compliance tooling for German SMEs}
}
```

---

## About KlarKI

KlarKI is an open-source, local-first EU AI Act + GDPR compliance auditor built for German SMEs.
Upload a policy document and receive a scored gap analysis against Articles 9–15 entirely on your own hardware.

**Key features:**
- Deterministic legal decision hierarchy (actor detection, Annex III applicability gate)
- Hybrid RAG retrieval (BM25 + ChromaDB vector + cross-encoder re-ranking)
- LangGraph multi-agent gap analysis (3-node per applicable article)
- Bilingual EN/DE support — all inference runs locally, no external API calls

[GitHub](https://github.com/s4nkar/KlarKI-EU-AI-Act-compliance-auditor)  |  [All KlarKI Models](https://huggingface.co/s4nkar)