README.md

---
language:
- it
- en
- de
- fr
- es
- multilingual
license:
- mit
datasets:
- xtreme
metrics:
- precision: 0.874
- recall: 0.88
- f1: 0.877
- accuracy: 0.943
inference:
  parameters:
    aggregation_strategy: first
---

# gunghio/xlm-roberta-base-finetuned-panx-ner

This model was trained starting from xlm-roberta-base on a subset of xtreme dataset.

`xtreme` datasets subsets used are: PAN-X.{lang}. Language used for training/validation are: italian, english, german, french and spanish.

Only 75% of the whole dataset was used.

## Intended uses & limitations

Fine-tuned model can be used for Named Entity Recognition in it, en, de, fr, and es.

## Training and evaluation data

Training dataset: [xtreme](https://huggingface.co/datasets/xtreme)

### Training results

It achieves the following results on the evaluation set:

- Precision: 0.8744154472771157
- Recall: 0.8791424269015351
- F1: 0.8767725659462058
- Accuracy: 0.9432040948504613

Details:

| Label   | Precision | Recall | F1-Score | Support |
|---------|-----------|--------|----------|---------|
| PER     | 0.922     | 0.908  | 0.915    | 26639   |
| LOC     | 0.880     | 0.906  | 0.892    | 37623   |
| ORG     | 0.821     | 0.816  | 0.818    | 28045   |
| Overall | 0.874     | 0.879  | 0.877    | 92307   |

## Usage

Set aggregation stragey according to [documentation](https://huggingface.co/docs/transformers/v4.18.0/en/main_classes/pipelines#transformers.TokenClassificationPipeline).

```python
from transformers import AutoTokenizer, AutoModelForTokenClassification
from transformers import pipeline

tokenizer = AutoTokenizer.from_pretrained("gunghio/xlm-roberta-base-finetuned-panx-ner")
model = AutoModelForTokenClassification.from_pretrained("gunghio/xlm-roberta-base-finetuned-panx-ner")

nlp = pipeline("ner", model=model, tokenizer=tokenizer, aggregation_strategy="first")
example = "My name is Wolfgang and I live in Berlin"

ner_results = nlp(example)
print(ner_results)
```