metadata

license: cc-by-nc-sa-4.0
datasets:
  - somosnlp/es-inclusive-language
language:
  - es
metrics:
  - bleurt
  - sacrebleu
pipeline_tag: text2text-generation
tags:
  - social

Model Card for Traductor Inclusivo

This model is a fine-tuned version of projecte-aina/aguila-7b on the dataset somosnlp/es-inclusive-language.

Languages are powerful tools to communicate ideas, but their use is not impartial. The selection of words carries inherent biases and reflects subjective perspectives. In some cases, language is wielded to enforce ideologies, marginalize certain groups, or promote specific political agendas. Spanish is not the exception to that. For instance, when we say “los alumnos” or “los ingenieros”, we are excluding women from those groups. Similarly, expressions such as “los gitanos” o “los musulmanes” perpetuate discrimination against these communities.

In response to these linguistic challenges, this model offers a way to construct inclusive alternatives in accordance with official guidelines on inclusive language from various Spanish speaking countries. Its purpose is to provide grammatically correct and inclusive solutions to situations where our language choices might otherwise be exclusive. By rectifying biases ingrained in language and fostering inclusivity, it combats discrimination, amplifies the visibility of marginalized groups, and contributes to the cultivation of a more inclusive and respectful society. This is a tool that contributes to the Sustainable Development Goals number five (Achieve gender equality and empower all women and girls) and ten (Reduce inequality within and among countries).

The model works in such a way that, given an input text, it returns the original text rewritten using inclusive language.

It achieves the following results on the validation set:

Loss: 0.6030

Model Details

Model Description

Developed by: Andrés Martínez Fernández-Salguero, Imanuel Rozenberg, Gaia Quintana Fleitas, Miguel López Pérez and Josué Sauca
Funded by: SomosNLP, HuggingFace
Model type: Language model, instruction tuned
Language(s): Spanish (es-ES, es-AR, es-MX, es-CR, es-CL)
License: cc-by-nc-sa-4.0
Fine-tuned from model: projecte-aina/aguila-7b
Dataset used: somosnlp/es-inclusive-language

Model Sources

Repository: https://github.com/Andresmfs/Traductor_inclusivo
Demo: https://huggingface.co./spaces/somosnlp/es-inclusive-language-demo
Video presentation: https://www.youtube.com/watch?v=7rrNGJIXEHU

Uses

Direct Use

The general uses of this model are adaptations of texts in Spanish to inclusive language.

It can be used mainly to adapt news, blogposts, emails and official documents among others.

Out-of-Scope Use

This model is specifically designed for translating Spanish texts to Spanish texts in inclusive language. Using the model for unrelated tasks is considered out of scope. This model can not be used with commercial purposes, it is intended for research or educational purposes only.

Bias, Risks, and Limitations

Model has not been trained on long-complex texts.
Model has been trained mostly with sentences where the terms to be modified are at the beginning of the sentence.
Model returns only one translation option when several might also be adequate.
Possible small information omission on translation.
Possible forced use of the term "personas".
Model does not detect or modify hate speech.
Model has been trained on data mainly based on Spanish Inclusive Language Guidelines and may inherit any bias comming from the guidelines and institutions behind them. They are official and updated guidelines that should not contain strong biases.
Model may not work propperly on translation difficulties aside the list of difficulties present on es-inclusive-language dataset
Other biases coming from the train dataset es-inclusive-language dataset should be taken into account.

Recommendations

Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model.

How to Get Started with the Model

Use the code below to get started with the model in 16-bits.

from transformers import AutoTokenizer
from transformers import AutoModelForCausalLM
import torch

# Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained('somosnlp/', trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained('somosnlp/', trust_remote_code=True,
    quantization_config=bnb_config,
    device_map="auto")

# generation_config
generation_config = model.generation_config
generation_config.max_new_tokens = 100
generation_config.temperature = 0.7
generation_config.top_p = 0.7
generation_config.num_return_sequences = 1
generation_config.pad_token_id = tokenizer.eos_token_id
generation_config.eos_token_id = tokenizer.eos_token_id

# Define inference function
def translate_es_inclusivo(exclusive_text): 
    
    # generate input prompt
    eval_prompt = f"""Reescribe el siguiente texto utilizando lenguaje inclusivo.\n
      Texto: {exclusive_text}\n
      Texto en lenguaje inclusivo:"""
    
    # tokenize input
    model_input = tokenizer(eval_prompt, return_tensors="pt").to(model.device)
    
    # set max_new_tokens if necessary
    if len(model_input['input_ids'][0]) > 80:
        model.generation_config.max_new_tokens = len(model_input['input_ids'][0]) + 0.2 * len(model_input['input_ids'][0])
    
    # get length of encoded prompt
    prompt_token_len = len(model_input['input_ids'][0])
        
    # generate and decode
    with torch.no_grad():
        inclusive_text = tokenizer.decode(model.generate(**model_input, generation_config=generation_config)[0][prompt_token_len:], 
                                          skip_special_tokens=True)                                                                        
    
    return inclusive_text

##########

input_text = 'Los alumnos atienden a sus profesores'

print(translate_es_inclusivo(input_text))

As it is a heavy model, you may want to use it in 4-bits:

from transformers import AutoTokenizer
from transformers import AutoModelForCausalLM
from transformers import BitsAndBytesConfig
import torch


## Load model in 4bits
# bnb_configuration
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type='nf4',
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=False)


# model
model = AutoModelForCausalLM.from_pretrained('somosnlp/', trust_remote_code=True,
    quantization_config=bnb_config,
    device_map="auto")

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained('somosnlp/', trust_remote_code=True)

# generation_config
generation_config = model.generation_config
generation_config.max_new_tokens = 100
generation_config.temperature = 0.7
generation_config.top_p = 0.7
generation_config.num_return_sequences = 1
generation_config.pad_token_id = tokenizer.eos_token_id
generation_config.eos_token_id = tokenizer.eos_token_id

# Define inference function
def translate_es_inclusivo(exclusive_text): 
    
    # generate input prompt
    eval_prompt = f"""Reescribe el siguiente texto utilizando lenguaje inclusivo.\n
      Texto: {exclusive_text}\n
      Texto en lenguaje inclusivo:"""
    
    # tokenize input
    model_input = tokenizer(eval_prompt, return_tensors="pt").to(model.device)
    
    # set max_new_tokens if necessary
    if len(model_input['input_ids'][0]) > 80:
        model.generation_config.max_new_tokens = len(model_input['input_ids'][0]) + 0.2 * len(model_input['input_ids'][0])
    
    # get length of encoded prompt
    prompt_token_len = len(model_input['input_ids'][0])
        
    # generate and decode
    with torch.no_grad():
        inclusive_text = tokenizer.decode(model.generate(**model_input, generation_config=generation_config)[0][prompt_token_len:], 
                                          skip_special_tokens=True)                                                                        
    
    return inclusive_text

##########

input_text = 'Los alumnos atienden a sus profesores'

print(translate_es_inclusivo(input_text))

Training Details

Training Data

Train, validation and test data splits can be found in somosnlp/es-inclusive-language

Training Procedure

For training we used QLoRA technique in 4-bits and rank 8

Find the training script here

Training Hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0001
train_batch_size: 8
eval_batch_size: 8
seed: 42
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
num_epochs: 10
Training regime: fp16 mixed precision

Speeds, Sizes, Times

The model was trained in 10 epochs with a total duration of 2hours and 54 minutes.

Training Loss	Epoch	Step	Validation Loss
No log	1.0	402	0.8020
1.0274	2.0	804	0.7019
0.6745	3.0	1206	0.6515
0.5826	4.0	1608	0.6236
0.5104	5.0	2010	0.6161
0.5104	6.0	2412	0.6149
0.4579	7.0	2814	0.6030
0.4255	8.0	3216	0.6151
0.3898	9.0	3618	0.6209
0.3771	10.0	4020	0.6292

Evaluation

Testing Data, Factors & Metrics

Testing Data

Here you can find the validation set used during training. Here you can find the test set used for evaluating model errors.

Metrics

For test evaluation it has been used a weighted harmonic mean of metrics bleurt (60%) and Sacrebleu (40%).

In Sacrebleu metric grammatical correctness carries high weight compared to the actual words used, whereas in Bleurt metric the actual words used have higher weight over grammatical correctness. Combining both metrics, we account for a grammatically correct prediction together with the use of the required specific words.

Results

On this notebook you can find the results of the test evaluation.

We get an average score of 68.4 (measured with the above described metric). Due to the existence of equivalent language formulas (these are inclusive language formulas that can be used indistinctly and the choice of a formula over the other is rather a stylistic decision than a language correctness decision) it is possible to argue that the real score of the model is higher.

Environmental Impact

Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).

Hardware Type: Nvidia T4 medium (8 vCPU, 30 Gb RAM, 16 Gb VRAM)
Hours used: 3 hours
Cloud Provider: Google Cloud Platform
Compute Region: europe-west
Carbon Emitted: 0.13 kg CO2 eq.

Technical Specifications

Model Architecture and Objective

The base model is projecte-aina/aguila-7b finetuned in 4-bit.

Compute Infrastructure

Hardware

Hardware used was Nvidia T4 medium (8 vCPU, 30 Gb RAM, 16 Gb VRAM) funded by Hugging Face

Software

Transformers 4.30.0
Pytorch 2.2.2+cu121
Datasets 2.18.0
Tokenizers 0.13.3
Peft

License

Creative Commons (cc-by-nc-sa-4.0) This kind of license is inherited from dataset used for training.

Citation

BibTeX:

@software{AIGMJ2024TraductorInclusivo,
  author = {Andrés Martínez Fernández-Salguero, Imanuel Rozenberg, Gaia Quintana Fleitas, Miguel López Pérez, Josué Sauca},
  title = {TraductorInclusivo},
  month = April,
  year = 2024,
  url = {https://huggingface.co./somosnlp/es-inclusivo-translator}
}

AIGMJ2024TraductorInclusivo
author: Andrés Martínez Fernández-Salguero, Imanuel Rozenberg, Gaia Quintana Fleitas, Miguel López Pérez, Josué Sauca
title: Traductor Inclusivo
year: 2024
url: https://huggingface.co./somosnlp/es-inclusivo-translator

More Information

This project was developed during the Hackathon #Somos600M organized by SomosNLP. The model was trained using GPUs sponsored by HuggingFace.

Team:

somosnlp
/

es-inclusivo-translator