---
datasets:
- hackathon-somos-nlp-2023/DiagTrast
language:
- es
metrics:
- accuracy
license: mit
inference: false
---

# Model Card for "DiagTrast-Berto"

This model is a fine-tuned version of [dccuchile/bert-base-spanish-wwm-cased](https://huggingface.co/dccuchile/bert-base-spanish-wwm-cased), which is a BERT model trained on a big Spanish corpus.

DiagTrast-Berto was trained with [hackathon-somos-nlp-2023/DiagTrast](https://huggingface.co/datasets/hackathon-somos-nlp-2023/DiagTrast) dataset to classify statements with each of the 5 selected mental disorders of the DSM-5. While this task is classically approached with neural network-based models, the goal of implementing a transformer model is that instead of basing the classification criteria on keyword search, it is expected to understand natural language.

## Uses

The model can be used to classify statements written by professionals who have detected unusual behaviors or characteristics in their patients that would indicate the presence of a mental disorder; at the moment it only provides support for five of the disorders described in the DSM-5. It should be noted that the model aims to identify the predominant disorder, so it would be part of the professional's job to group the symptoms before entering them into the model for cases in which multiple disorders are presumed to be present at the same time.

### Direct Use

DiagTrast-Berto is already a fine-tuned model so it could be used directly to classify the statements.

### Out-of-Scope Use

This model should not be used as a replacement for a mental health professional because it is always necessary that each situation be evaluated responsibly and using all human intellectual capacity. Initially this model is designed as an auxiliary tool to facilitate the use of the DSM-5 by health professionals.

## Bias, Risks, and Limitations

The main limitation of the model is that it is restricted to the identification of only 5 of the DSM-5 disorders.

Also, the model will always match a statement with a disorder since there was not a 'non-disorder' label in the dataset.

## How to Get Started with the Model

Use the code below to get started with the model.

```python
>>> from transformers import pipeline
>>> classifier = pipeline("text-classification", model='hackathon-somos-nlp-2023/DiagTrast-Berto')
>>> text = ["Tiene pocas habilidades sociales, ignora normas de convivencia"]
>>> classifier.predict(text)
[{'label': 'Trastornos de la personalidad antisocial',
  'score': 0.9772088527679443}]
```

## Training Details

### Training Data

We use the [hackathon-somos-nlp-2023/DiagTrast](https://huggingface.co/datasets/hackathon-somos-nlp-2023/DiagTrast) dataset, it was split with 90% of records for the training set and 10% for the test set using the 'datasets' library of hugging face.

### Training Procedure 

We use HuggingFace's Transformers library to load [BERTO](https://huggingface.co/dccuchile/bert-base-spanish-wwm-cased) checkpoint and fine-tune the model.


#### Training Hyperparameters

We use the default ones.

## Evaluation

The valuation dataset consists of 134 arbitrarily selected examples, so labels may not be in the same proportion. We use 'Accuracy' as our metric, achieving a 97% accuracy after 3 epochs. 

## Environmental Impact

<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->

Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).

- **Hardware Type:** Tesla T4 
- **Hours used:** 0.09 hours
- **Cloud Provider:** Google
- **Compute Region:** Spain
- **Carbon Emitted:** 0.005 kg C02
  
## Team members

- [Alberto Martín Garrido](https://huggingface.co/Stremie)
- [Edgar Mencia](https://huggingface.co/edmenciab)
- [Miguel Ángel Solís Orozco](https://huggingface.co/homosapienssapiens)
- [Jose Carlos Vílchez Villegas](https://huggingface.co/JCarlos)