---
dataset_info:
  features:
  - name: tweet_id
    dtype: string
  - name: tokens
    sequence: string
  - name: ner_tags
    sequence:
      class_label:
        names:
          '0': B-PROFESION
          '1': I-PROFESION
          '2': O
  splits:
  - name: train
    num_bytes: 1835289
    num_examples: 2786
  - name: validation
    num_bytes: 614287
    num_examples: 999
  - name: test
    num_bytes: 623453
    num_examples: 1001
  download_size: 805537
  dataset_size: 3073029
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
---

# NER-PROFESION Dataset (Based on PROFNER)

This dataset is built upon the original **[PROFNER](https://zenodo.org/records/4563995)** corpus, which focuses on Named Entity Recognition (NER) of **professions in Spanish**.

To ensure consistency and simplicity, the original fine-grained entity tags have been **normalized** into the following three labels:

- `B-PROFESION`: beginning of a profession entity
- `I-PROFESION`: continuation of a profession entity
- `O`: outside any named entity

---

## 📦 Dataset Structure

The dataset is provided as a Hugging Face `DatasetDict` with the following splits:

| Split        | Description |
|--------------|-------------|
| `train`      | Balanced: approximately 50% of the documents contain at least one `B-PROFESION` tag, while the other 50% contain no profession entities. |
| `validation` | Balanced with respect to the number of documents containing profession entities, matching the `test` split. |
| `test`       | Also balanced: half of the documents contain at least one `B-PROFESION`, half contain none. |

---

## ⚙️ Dataset Generation

Two utility functions were used to prepare the dataset:

### `procesar_training_set_balanceado`

This function loads the original training data and performs the following steps:

1. Groups the data by document ID.
2. Splits documents into two groups: those containing at least one `B-` label (indicating a profession) and those without.
3. Selects an equal number of documents from both groups to ensure a 50/50 balance between positive and negative samples.
4. Converts the documents into CoNLL-style format: one token-label pair per line, separated by empty lines between documents.

### `procesar_dev_test_balanceado`

This function splits the original development set into two subsets:

1. Documents are separated into two groups:
   - With at least one `B-` tag
   - Without any `B-` tag
2. Each group is split equally into two halves:
   - Half goes into the `validation` set
   - Half goes into the `test` set
3. This guarantees that both `validation` and `test` are balanced with respect to the presence of annotated profession entities.

The selected document IDs are stored separately (`dev_ids.txt` and `test_ids.txt`) for reproducibility.

---

## 🧾 Format

Each instance is a dictionary with:

- `tokens`: list of tokenized words
- `ner_tags`: list of integer-encoded entity labels

The `ner_tags` feature uses Hugging Face's `ClassLabel`, which maps integers to string labels:

```python
label_list = ["B-PROFESION", "I-PROFESION", "O"]
```

Example:

```python
{
  "tokens": ["Nuestros", "colaboradores", "y", "conductores"],
  "ner_tags": [0, 1, 2, 0]  # Using ClassLabel -> ["B-PROFESION", "I-PROFESION", "O", "B-PROFESION"]
}
```

---

## 🔁 Label Mappings

The dataset includes a `label_mappings.json` file with:

```json
{
  "label_list": ["B-PROFESION", "I-PROFESION", "O"],
  "label2id": {
    "B-PROFESION": 0,
    "I-PROFESION": 1,
    "O": 2
  },
  "id2label": {
    "0": "B-PROFESION",
    "1": "I-PROFESION",
    "2": "O"
  }
}
```

These can be loaded to configure a model or to interpret predictions.

---

## 📥 Loading the Dataset

```python
from datasets import load_dataset

dataset = load_dataset("luisgasco/profner_ner_master")
tokens = dataset["train"][0]["tokens"]
tags = dataset["train"].features["ner_tags"].feature.int2str(dataset["train"][0]["ner_tags"])

print(list(zip(tokens, tags)))
```


---

## 🔗 Original Data Source

This dataset is based on the **PROFNER** corpus:
- Zenodo: [LINK](https://zenodo.org/records/4563995)
- [Original task website](https://temu.bsc.es/smm4h-spanish/)

---

## ✍️ Author

This version has been processed and curated by [Luis Gasco](https://huggingface.co/luisgasco), based on the PROFNER dataset, for educational purposes.

---