README.md

---
license: cc-by-sa-4.0
task_categories:
  - token-classification
language:
  - de
tags:
  - ner
  - named-entity-recognition
  - multilingual
  - sdvm
  - refined
pretty_name: SDVM Refined PAN-X.de (XTREME NER)
dataset_info:
  features:
    - name: tokens
      sequence: string
    - name: ner_tags
      sequence:
        class_label:
          names:
            '0': O
            '1': B-PER
            '2': I-PER
            '3': B-ORG
            '4': I-ORG
            '5': B-LOC
            '6': I-LOC
    - name: tokens_refined
      sequence: string
    - name: ner_tags_refined
      sequence:
        class_label:
          names:
            '0': O
            '1': B-PER
            '2': I-PER
            '3': B-ORG
            '4': I-ORG
            '5': B-LOC
            '6': I-LOC
    - name: langs
      sequence: string
  splits:
    - name: train
      num_examples: 20000
    - name: validation
      num_examples: 10000
    - name: test
      num_examples: 10000
---

# SDVM Refined PAN-X.de (XTREME NER)

This is an enhanced version of the PAN-X German NER dataset from the [XTREME benchmark](https://huggingface.co/datasets/google/xtreme) (config: PAN-X.de), refined by SDVM (Synthetic Data Vending Machine) to improve token quality for Named Entity Recognition.

## Refinement

The original PAN-X.de dataset is derived from WikiANN/Wikipedia and contains various artifacts from the extraction process. SDVM applied automated token refinement that removes Wikipedia markup noise while preserving entity annotations:

- **Markup token removal**: Removed ~8.5% of tokens that were Wikipedia formatting artifacts (bold markers, quote marks, section headers, template brackets, etc.)
- **Redirect cleanup**: Removed `WEITERLEITUNG` (German Wikipedia redirect markers)
- **Embedded markup stripping**: Cleaned tokens containing partial markup (e.g., `Friedrichsaue` with trailing brackets)
- **Tag continuity repair**: Automatically fixed B-/I- tag sequences after token removal to maintain valid IOB2 format

## Dataset Structure

Each example contains both the original and refined versions:

| Column | Description |
|--------|-------------|
| `tokens` | Original token sequence |
| `ner_tags` | Original NER tags (IOB2 format) |
| `tokens_refined` | Cleaned token sequence |
| `ner_tags_refined` | Adjusted NER tags for refined tokens |
| `langs` | Language tags |

### NER Labels

| ID | Tag |
|----|-----|
| 0 | O |
| 1 | B-PER |
| 2 | I-PER |
| 3 | B-ORG |
| 4 | I-ORG |
| 5 | B-LOC |
| 6 | I-LOC |

## Example

**Original**: `['**', "''", 'Lou', 'Salome', "''", '.']` with tags `[O, O, B-PER, I-PER, O, O]`

**Refined**: `['Lou', 'Salome', '.']` with tags `[B-PER, I-PER, O]`

## Usage

```python
from datasets import load_dataset

# Load the refined dataset
ds = load_dataset("SDVM/xtreme-PAN-X.de")

# Use refined columns for training
train_tokens = ds["train"]["tokens_refined"]
train_tags = ds["train"]["ner_tags_refined"]

# Or use original columns for comparison
orig_tokens = ds["train"]["tokens"]
orig_tags = ds["train"]["ner_tags"]
```

## Source

- **Original dataset**: [google/xtreme](https://huggingface.co/datasets/google/xtreme) (PAN-X.de config)
- **Reference**: Chapter 4 of [Natural Language Processing with Transformers](https://github.com/nlp-with-transformers/notebooks/blob/main/04_multilingual-ner.ipynb)
- **Refinement by**: [SDVM](https://huggingface.co/SDVM)