Model Card for bertin_base_climate_detection_spa_v2

README Spanish Version: README_ES

Model Illustration

This model is a fine-tuning version of the model: bertin-project/bertin-roberta-base-spanish using the dataset somosnlp/spa_climate_detection. The model is focused on the identification of texts on topics related to climate change and sustainability. This project was based on the English version of climatebert/distilroberta-base-climate-detector.

The motivation of the project was to create a repository in Spanish on information or resources on topics such as: climate change, sustainability, global warming, energy, etc; the idea is to give visibility to solutions, examples of good environmental practices or news that help us to combat the effects of climate change; in a way similar to what the project Drawdown does but providing examples of solutions or new research on each topic. To achieve this In order to achieve this objective, we consider that the identification of texts that speak about these topics is the first step. Some of the direct applications are: classification of papers and scientific publications, news, opinions.

Future steps:

We intend to create an advanced model that classifies texts related to climate change based on sectors (token classification), for example: classify based on electricity, agriculture, industry, transport, etc.
Publish a sector-based dataset.
Realize a Q/A model that can provide relevant information to the user on the topic of climate change.

Model Details

Model Description

Developed by: Gerardo Huerta Gabriela Zuñiga
Funded by: SomosNLP, HuggingFace
Model type: Language model, instruction tuned, text classification
Language(s): es-ES, es-PE
License: cc-by-nc-sa-4.0
Fine-tuned from model: bertin-project/bertin-roberta-base-spanish
Dataset used: somosnlp/spa_climate_detection

Model resurces:

Repository: somosnlp/bertin_base_climate_detection_spa
Demo: identificacion de textos sobre cambio climatico y sustentabilidad
Video presentation: Proyecto BERTIN-ClimID

Uses

Direct Use

News classification: With this model it is possible to classify news headlines related to the areas of climate change.
Paper classification: The identification of scientific texts that disclose solutions and/or effects of climate change. For this use, the abstract of each paper can be used for identification.

Indirect Use

For the creation of information repositories regarding climate issues.
This model can serve as a basis for creating new classification systems for climate solutions to disseminate new efforts to combat climate change in different sectors.
Creation of new datasets that address the issue.

Out-of-Scope Use

The use for text classification of unverifiable or unreliable sources and their dissemination, e.g., fake news or disinformation.

Bias, Risks, and Limitations

No specific studies on biases and limitations have been carried out at this point, however, we make the following points based on previous experience and model tests:

It inherits the biases and limitations of the base model with which it was trained, for more details see: BERTIN: Efficient Pre-Training of a Spanish Language Model using Perplexity Sampling. However, they are not so obvious to find because of the type of task in which the model is being implemented, such as text classification.
Direct biases such as the majority use of high-level language in the dataset due to the use of texts extracted from news, legal documentation of companies that can complicate the identification of texts with low-level language (e.g. colloquial). To mitigate these biases, diverse opinions on climate change issues extracted from sources such as social networks were included in the dataset, in addition to a rebalancing of the labels.
The dataset inherits other limitations such as: the model loses performance on short texts, this is due to the fact that most of the texts used in the dataset have a long length between 200 - 500 words. Again, we tried to mitigate these limitations by including short texts.

Recommendations

As we have mentioned, the model tends to lower performance in short texts, so it is advisable to establish a selection criterion for long texts whose subject matter needs to be identified.

How to Get Started with the Model

## Asumiendo tener instalados transformers, torch
from transformers import AutoModelForSequenceClassification
import torch
from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("somosnlp/bertin_base_climate_detection_spa")
model = AutoModelForSequenceClassification.from_pretrained("somosnlp/bertin_base_climate_detection_spa")

# Traduccion del label 
id2label = {0: "NEGATIVE", 1: "POSITIVE"}
label2id = {"NEGATIVE": 0, "POSITIVE": 1}

# Funcion de inferencia
def inference_fun(Texto):
    inputs = tokenizer(Texto, return_tensors="pt")
    with torch.no_grad():
        logits = model(**inputs).logits
    predicted_class_id = logits.argmax().item()
    output_tag = model.config.id2label[predicted_class_id]
    return output_tag

input_text = "El uso excesivo de fertilizantes nitrogenados -un fenómeno frecuente en la agricultura- da lugar a la producción de óxido nitroso, un potente gas de efecto invernadero. Un uso más juicioso de los fertilizantes puede frenar estas emisiones y reducir la producción de fertilizantes, que consume mucha energía."

print(inference_fun(input_text))

Training Details

Training Data

The training data were obtained from the dataset somosnlp/spa_climate_detection. The training data represent about 79% of the total data in the dataset.

The labels are represented as follows:

Labels 1s

1000 - data on paragraphs extracted from company reports on the subject.

600 - data on various opinions, mostly short texts.

Labels 0s

300 - data on paragraphs extracted from business reports not related to the subject.

500 - data on news on various topics unrelated to the subject.

500 - data on opinions on various topics unrelated to the subject.

Training Procedure

You can check our Google Colab to review the training procedure we take: Colab Entrenamiento The accelerate configuration is as follows: In which compute environment are you running?: 0 Which type of machine are you using?: No distributed training Do you want to run your training on CPU only (even if a GPU / Apple Silicon / Ascend NPU device is available)? [yes/NO]:NO Do you wish to optimize your script with torch dynamo?[yes/NO]:NO Do you want to use DeepSpeed? [yes/NO]: NO What GPU(s) (by id) should be used for training on this machine as a comma-seperated list? [all]:all Do you wish to use FP16 or BF16 (mixed precision)?: no

Training Hyperparameters

The following hyperparameters were used during training:

learning_rate: 2e-05
train_batch_size: 16
eval_batch_size: 16
seed: 42
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
num_epochs: 2

Speeds, Sizes, Times

The model was trained in 2 epochs with a total training duration of 14.22 minutes, 'train_runtime': 853.6759. Additional information: No mixed precision (FP16 or BF16) was used.

Resultados del entrenamiento:

Training Loss	Epoch	Step	Validation Loss	Accuracy
No log	1.0	182	0.1964	0.9551
No log	2.0	364	0.1592	0.9705

Evaluation

Testing Data, Factors & Metrics

Testing Data

The assessment data were obtained from the dataset somosnlp/spa_climate_detection. The assessment data represent about 21% of the total data in the dataset. The labels are represented as follows:

Labels 1s

320 - data on paragraphs extracted from company reports on the subject.

160 - data on various opinions, mostly short texts.

Labels 0s

80 - data on paragraphs extracted from business reports not related to the subject.

120 - data on news on various topics unrelated to the subject.

100 - data on opinions on various topics unrelated to the subject.

Model reached the following results on evaluation dataset:

Loss: 0.1592
Accuracy: 0.9705

Metrics

The metric was precision.

Results

Look at the Inference section of Colab: entrenamiento_del_modelo

Accuracy 0.95 Precision 0.916 Recall 0.99 F1 score 0.951

Environmental Impact

Using the tool ML CO2 IMPACT we estimate the following environmental impact due to training:

Type of hardware: T4
Total Hours for iterations and tests: 4 horas
Cloud provider Google Cloud (colab)
Computational region us-east
Carbon footprint 0.1kg CO2

Technical Specifications

Software

Transformers 4.39.3
Pytorch 2.2.1+cu121
Datasets 2.18.0
Tokenizers 0.15.2

Hardware

GPU equivalent to T4
For reference, the model was trained on the free version of Google Colab

License

cc-by-nc-sa-4.0 Due to inheritance of the data used in the dataset.

Citation

BibTeX:

@software{BERTIN-ClimID,
  author = {Gerardo Huerta, Gabriela Zuñiga},
  title = {BERTIN-ClimID: BERTIN-Base Climate-related text Identification},
  month = Abril,
  year = 2024,
  url = {https://huggingface.co/somosnlp/bertin_base_climate_detection_spa}
}

More Information

This project was developed during the Hackathon #Somos600M organized by SomosNLP. We thank all event organizers and sponsors for their support during the event.

Team:

somosnlp
/

bertin_base_climate_detection_spa