|
--- |
|
license: mit |
|
datasets: |
|
- BrunoGR/HEAR-Hispanic_Emotional_Accompaniment_Responses |
|
- >- |
|
BrunoGR/HRECPW-Hispanic_Responses_for_Emotional_Classification_based_on_Plutchik_Wheel |
|
language: |
|
- es |
|
tags: |
|
- Emotional |
|
- Emotional Support |
|
- Emotional Accompaniment |
|
- chatbot |
|
library_name: transformers |
|
pipeline_tag: text-generation |
|
--- |
|
|
|
# Sólo Escúchame: Spanish Emotional Accompaniment Chatbot 💬🤖 |
|
|
|
Sólo Escúchame is an open-source Spanish emotional assistance chatbot designed to provide psychological support. It is built upon the LLaMA-2-7b-Chat model and fine-tuned using the HEAR (Hispanic Emotional Accompaniment Responses) dataset. |
|
|
|
## Overview |
|
|
|
Mental health issues have been rapidly increasing, with suicide being the fourth leading cause of death among individuals aged 15 to 29 in 2019, according to the World Health Organization (WHO). Sólo Escúchame aims to address this urgent need by offering a supplementary tool for psychological support, especially for Spanish speakers who may not have immediate access to professional help. |
|
|
|
## Features |
|
|
|
- **Emotional Assistance**: Provides empathetic and supportive responses to users' emotional situations. |
|
- **HEAR Dataset**: Trained on a specialized dataset for emotional accompaniment, compiled from multiple English sources and translated into Spanish. |
|
- **Open-Source**: Available for public use and contribution, facilitating reproducibility and further research. |
|
- **CPU Efficient**: Runs efficiently on CPUs, making it accessible to a wider audience. |
|
|
|
## Model |
|
|
|
Sólo Escúchame is a fine-tuned version of the LLaMA-2-7b-Chat model. It utilizes the Rotary Positional Embedding (RoPE) and Grouped-Query Attention (GQA) techniques to enhance context length and model performance. The model has been quantized to 2, 4, and 8 bits to ensure accessibility. |
|
|
|
## Training |
|
|
|
The model was trained using LoRA (Low Rank Adaptation) on the HEAR dataset. The training parameters were optimized for performance and efficiency. |
|
|
|
## Dataset |
|
|
|
### Hispanic Emotion Recognition Based on Plutchik’s Wheel (HRECPW) Dataset |
|
|
|
- **Source**: Translated from diverse English sources including TweetEval, DailyDialog, HappyDB, and survey responses. |
|
- **Classes**: 11 emotion classes - affection, happiness, admiration, anger, sadness, optimism, hate, surprise, fear, calm, and disgust. |
|
- **Size**: 121,000 training examples, 2,200 validation examples, and 1,320 test examples. |
|
|
|
### Hispanic Emotional Accompaniment Responses (HEAR) Dataset |
|
|
|
- **Purpose**: Used to train the Sólo Escúchame model for generating empathetic and suitable responses. |
|
- **Size**: 41,481 training examples, 2,200 validation examples, and 1,320 test examples. |
|
|
|
## Evaluation |
|
|
|
The model's performance was evaluated using two main criteria: |
|
|
|
### Active Listening Technique |
|
|
|
| Evaluation trait | GPT-3.5 | LLaMA-2-7b-Chat | Mixtral8x7b | GPT-2-124M | Solo Escúchame | |
|
|:----------------------------:|:-------:|:---------------:|:-----------:|:----------:|:--------------:| |
|
| **Contextual Attention** | 1256 | 1260 | 1277 | 462 | **1240** | |
|
| **Clarifying Questions** | 776 | 718 | 531 | 199 | **913** | |
|
| **Deeper Conversation** | 1215 | 1240 | 1185 | 470 | **1254** | |
|
| **Absence of Judgment** | 1292 | 1278 | 1299 | 517 | **1300** | |
|
| **Demonstration of Empathy** | 1246 | 1274 | 1287 | 502 | **1278** | |
|
|
|
### Socratic Method |
|
|
|
| Evaluation trait | GPT-3.5 | LLaMA-2-7b-Chat | Mixtral8x7b | GPT-2-124M | Solo Escúchame | |
|
|:-------------------------------------------:|:-------:|:---------------:|:-----------:|:----------:|:--------------:| |
|
| **Use of Inductive Questions** | 1077 | 1033 | 872 | 502 | **1224** | |
|
| **Non-Imposition of Ideas** | 1236 | 1170 | 1200 | 536 | **1299** | |
|
| **Expansion and Construction of Knowledge** | 1031 | 1071 | 972 | 473 | **1245** | |
|
| **Generation of Cognitive Dissonance** | 45 | 36 | 34 | 16 | **69** | |
|
| **Guided Discovery** | 1089 | 1076 | 988 | 498 | **1253** | |
|
|
|
### Final Scores for Psychological Accompaniment Evaluation |
|
|
|
| **Model** | **Active Listening** | **Socratic Method** | |
|
|:-------------------------:|:--------------------:|:-------------------:| |
|
| GPT2-124M | 32.57 | 30.68 | |
|
| Mixtral 8x7b | 84.52 | 61.60 | |
|
| LLaMA-2-7b-Chat | 87.42 | 66.45 | |
|
| GPT-3.5 | 87.62 | 67.84 | |
|
| **Sólo Escúchame (ours)** | **90.67** | **77.12** | |
|
|
|
*Table: Final Scores for Psychological Accompaniment Evaluation in Language Models (LMs)* |
|
|
|
## Usage |
|
|
|
The Sólo Escúchame model and datasets are publicly available on Hugging Face: |
|
|
|
- **Model**: [Sólo Escúchame](https://huggingface.co/BrunoGR/Just_HEAR_Me) |
|
- **Datasets**: |
|
- [HRECPW Dataset](https://huggingface.co/datasets/BrunoGR/HRECPW-Hispanic_Responses_for_Emotional_Classification_based_on_Plutchik_Wheel) |
|
- [HEAR Dataset](https://huggingface.co/datasets/BrunoGR/HEAR-Hispanic_Emotional_Accompaniment_Responses) |
|
|
|
## Installation and Setup |
|
|
|
To use the Sólo Escúchame model, follow these steps: |
|
|
|
1. Clone the repository: `git clone https://github.com/BrunoGilRamirez/Just_HEAR_ME` |
|
2. Install the required dependencies: `pip install -r requirements.txt` |
|
3. Load the model and dataset from Hugging Face: `from transformers import AutoModelForCausalLM, AutoTokenizer` |
|
|
|
|
|
## License |
|
|
|
Sólo Escúchame is released under the MIT License. |
|
|
|
|
|
## Citation |
|
|
|
If you use Sólo Escúchame (Just_HEAR_Me) in your research, please cite the following paper: |
|
|
|
```bibtex |
|
@article{Gil2024, |
|
title={Sólo Escúchame: Spanish Emotional Accompaniment Chatbot}, |
|
author={Gil Ramírez, Bruno and López Espejel, Jessica and Santiago Díaz, María del Carmen and Rubín Linares, Gustavo Trinidad}, |
|
journal={arxiv}, |
|
year={2024} |
|
} |
|
``` |
|
|
|
## Contact |
|
|
|
For any questions or inquiries, please contact: |
|
|
|
- Bruno Gil Ramírez: [email protected] |
