---
license: mit
language:
- multilingual
tags:
- zero-shot-classification
- text-classification
- pytorch
metrics:
- accuracy
- f1-score
extra_gated_prompt: 'Our models are intended for academic use only. If you are not
  affiliated with an academic institution, we will reject your access request. In
  addition, we would like to know which of our models are being used and by who.

  If you use our models for your work or research, please cite this paper: Sebők,
  M., Máté, Á., Ring, O., Kovács, V., & Lehoczki, R. (2024). Leveraging Open Large
  Language Models for Multilingual Policy Topic Classification: The Babel Machine
  Approach. Social Science Computer Review, 0(0). https://doi.org/10.1177/08944393241259434'
extra_gated_fields:
  Name: text
  Country: country
  Institution: text
  E-mail: text
---
# xlm-roberta-large-social-cap-v3
## Model description
An `xlm-roberta-large` model finetuned on multilingual training data containing texts of the `social` domain labelled with [major topic codes](https://www.comparativeagendas.net/pages/master-codebook) from the [Comparative Agendas Project](https://www.comparativeagendas.net/).

## How to use the model
#### Loading and tokenizing input data
```python
import pandas as pd
import numpy as np
from datasets import Dataset
from transformers import (AutoModelForSequenceClassification, AutoTokenizer, 
                          Trainer, TrainingArguments)

CAP_NUM_DICT = {0: '1', 1: '2', 2: '3', 3: '4', 4: '5', 5: '6', 
6: '7', 7: '8', 8: '9', 9: '10', 10: '12', 11: '13', 12: '14', 
13: '15', 14: '16', 15: '17', 16: '18', 17: '19', 18: '20', 19: 
'21', 20: '23', 21: '999'}

tokenizer = AutoTokenizer.from_pretrained('xlm-roberta-large')
num_labels = len(CAP_NUM_DICT)

def tokenize_dataset(data : pd.DataFrame):
    tokenized = tokenizer(data["text"],
                          max_length=MAXLEN,
                          truncation=True,
                          padding="max_length")
    return tokenized

hg_data = Dataset.from_pandas(data)
dataset = hg_data.map(tokenize_dataset, batched=True, remove_columns=hg_data.column_names)
```

#### Inference using the Trainer class
```python
model = AutoModelForSequenceClassification.from_pretrained('poltextlab/xlm-roberta-large-social-cap-v3',
                                                           num_labels=22,
                                                           problem_type="multi_label_classification",
                                                           ignore_mismatched_sizes=True
                                                           )

training_args = TrainingArguments(
    output_dir='.',
    per_device_train_batch_size=8,
    per_device_eval_batch_size=8
)

trainer = Trainer(
    model=model,
    args=training_args
)

probs = trainer.predict(test_dataset=dataset).predictions
predicted = pd.DataFrame(np.argmax(probs, axis=1)).replace({0: CAP_NUM_DICT}).rename(
    columns={0: 'predicted'}).reset_index(drop=True)

```

### Fine-tuning procedure
`xlm-roberta-large-social-cap-v3` was fine-tuned using the Hugging Face Trainer class with the following hyperparameters:
```python
training_args = TrainingArguments(
    output_dir=f"../model/{model_dir}/tmp/",
    logging_dir=f"../logs/{model_dir}/",
    logging_strategy='epoch',
    num_train_epochs=10,
    per_device_train_batch_size=8,
    per_device_eval_batch_size=8,
    learning_rate=5e-06,
    seed=42,
    save_strategy='epoch',
    evaluation_strategy='epoch',
    save_total_limit=1,
    load_best_model_at_end=True
)
```
We also incorporated an EarlyStoppingCallback in the process with a patience of 2 epochs.

## Model performance
The model was evaluated on a test set of 3488 examples (10% of the available data).<br>
Model accuracy is **0.74**.
| label        |   precision |   recall |   f1-score |   support |
|:-------------|------------:|---------:|-----------:|----------:|
| 0            |        0.69 |     0.71 |       0.7  |       184 |
| 1            |        0.65 |     0.61 |       0.63 |       167 |
| 2            |        0.7  |     0.72 |       0.71 |       105 |
| 3            |        0.88 |     0.82 |       0.85 |        88 |
| 4            |        0.79 |     0.75 |       0.77 |       145 |
| 5            |        0.84 |     0.89 |       0.86 |       186 |
| 6            |        0.79 |     0.83 |       0.81 |       169 |
| 7            |        0.95 |     0.78 |       0.86 |        73 |
| 8            |        0.86 |     0.79 |       0.82 |       357 |
| 9            |        0.86 |     0.88 |       0.87 |       221 |
| 10           |        0.79 |     0.81 |       0.8  |       296 |
| 11           |        0.66 |     0.8  |       0.72 |       239 |
| 12           |        0.69 |     0.57 |       0.62 |        77 |
| 13           |        0.74 |     0.48 |       0.58 |        58 |
| 14           |        0.57 |     0.65 |       0.61 |        31 |
| 15           |        0.7  |     0.42 |       0.53 |        45 |
| 16           |        0    |     0    |       0    |         8 |
| 17           |        0.48 |     0.46 |       0.47 |        89 |
| 18           |        0.69 |     0.7  |       0.69 |       527 |
| 19           |        0.71 |     0.7  |       0.71 |       103 |
| 20           |        0.74 |     0.69 |       0.72 |       108 |
| 21           |        0.63 |     0.72 |       0.68 |       212 |
| macro avg    |        0.7  |     0.67 |       0.68 |      3488 |
| weighted avg |        0.74 |     0.74 |       0.74 |      3488 |

## Inference platform
This model is used by the [CAP Babel Machine](https://babel.poltextlab.com), an open-source and free natural language processing tool, designed to simplify and speed up projects for comparative research.  

## Cooperation
Model performance can be significantly improved by extending our training sets. We appreciate every submission of CAP-coded corpora (of any domain and language) at poltextlab{at}poltextlab{dot}com or by using the [CAP Babel Machine](https://babel.poltextlab.com).

## Debugging and issues
This architecture uses the `sentencepiece` tokenizer. In order to run the model before `transformers==4.27` you need to install it manually.

If you encounter a `RuntimeError` when loading the model using the `from_pretrained()` method, adding `ignore_mismatched_sizes=True` should solve the issue.