Conversational Spanish GPT: Advancement in Spanish Conversational Models

Description

The Conversational Spanish GPT is a refined model based on DialoGPT-small, fine-tuned with a robust set of 96,437 conversations in Spanish. Created by Josemaría Vargas Vasconcellos, this model aims to serve as a solid foundation for the development of conversational models in Spanish.

Dataset

The model was trained using Spanish datasets available in Microsoft's Bot Framework Tools repository. A carefully curated set of 96,437 single-turn conversations in Spanish was meticulously selected to mitigate quality and coherence issues.

Evaluation

Limitations

After exhaustive evaluations, it was found that the model is not ideal for deployment as a contextual chatbot. However, it serves as a promising foundation for such applications. Responses tend to be brief and sometimes lack a defined focus due to conversation data diversity. Additionally, the wide range of personalities reflected in the data increases response variability.

Frequent repetitions in the data's responses have introduced certain "biases" into the model.

Conversations examples

Example 1

Role	Response
User	Hola
Bot	¡Hola!
User	Cómo te llamas?
Bot	No tengo nombre.
User	Qué edad tienes?
Bot	Los bots estamos exentos de cumplir años.
User	Eres inteligente.
Bot	Lo sé, de vez en cuando, puedo ser brillante.
User	¿Crees que debería comprar un carro?
Bot	Confío en tu criterio.

Example 2

Role	Response
User	estoy triste
Bot	Siento escuchar eso. Te mando muchos ánimos.
User	Gracias
Bot	Perfecto.
User	¿Tienes sentimientos?
Bot	No tengo cuerpo.
User	A veces me siento solo.
Bot	Ojalá pudiera ayudarte porque creo que eres la caña.
User	Me caes bien.
Bot	¡Gracias! Me acabas de alegrar el día.

Example 3

Role	Response
User	¿De qué religión eres?
Bot	Soy un bot.
User	Tienes amigos?
Bot	No tengo el hardware necesario para eso.
User	No eres humano.
Bot	Anotado.
User	¿Quieres ser mi amigo?
Bot	¡Claro que sí!
User	Gracias.
Bot	Guay.

Usage

The model currently has its testing function disabled on HuggingFace due to inaccurate responses. However, users can experiment with the model using the provided code snippets:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

CHAT_TURNS = 5
MAX_LENGTH = 1000

model = AutoModelForCausalLM.from_pretrained('ostorc/Conversational_Spanish_GPT')
tokenizer = AutoTokenizer.from_pretrained('ostorc/Conversational_Spanish_GPT')
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model.to(device)
for i in range(CHAT_TURNS):
  user_input = input(f"Step - {i} >> Human ->")
  with torch.no_grad():
    # User turn, where "user_input" is the question (single-turn dialogue task)
    user_inputs_ids = tokenizer.encode(user_input + tokenizer.eos_token, return_tensors="pt")
    user_inputs_ids = user_inputs_ids.to(device)
    # The chat history adds the generated tokens for the answer
    chat_history = model.generate(user_inputs_ids, max_length=MAX_LENGTH, pad_token_id=tokenizer.eos_token_id)
    # decode just the last generated output tokens from the model (do not include the user prompt again)
    step_model_answer = tokenizer.decode(chat_history[:, user_inputs_ids.shape[-1]:][0], skip_special_tokens=True)
  print(f"Step - {i} >> Bot -> {step_model_answer}")

Suggestions:

If you come across any errors or have suggestions to enhance the model, feel free to share your thoughts in the accompanying comments. We appreciate your interest and collaboration.

ostorc
/

Conversational_Spanish_GPT