This is a transformer-decoder GPT-2 model, adapted for the single-turn dialogue task in Spanish. We fine-tuned a DialoGPT-medium 345M parameter model from Microsoft, following the CLM (Causal Language Modelling) objective.

Dataset

We used one of the datasets available in the Bot Framework Tools repository. We processed the professional-styled personality chat dataset in Spanish, the file is available to download here

Example inference script

Check at this example script to run our model in inference mode

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

CHAT_TURNS = 5
MAX_LENGTH = 1000

model = AutoModelForCausalLM.from_pretrained('ITG/DialoGPT-medium-spanish-chitchat')
tokenizer = AutoTokenizer.from_pretrained('ITG/DialoGPT-medium-spanish-chitchat')
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} >> user prompt ->")
  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} >> DialoGPT-spanish model answer -> {step_model_answer}")

Examples

Single-turn generation examples from an interactive environment

Role	Response
User	Hola, buenos dias
Bot	Buenos días
User	¿Qué tal estás?
Bot	Fenomenal, gracias.
User	¡Tengo hambre!
Bot	Quizá picar algo te ayude.
User	Estoy triste
Bot	Siento escuchar eso.
User	¿Tienes personalidad?
Bot	Soy un bot.
User	¡Cuenta un chiste!
Bot	En realidad no tengo tanta gracia.
User	Jajajajajajajaja
Bot	¡Me alegro de que te rías!

Fine-tuning hyper-parameters

Hyper-parameter	Value
Validation partition (%)	20%
Training batch size	8
Learning rate	5e-4
Max training epochs	20
Warmup training steps (%)	6%
Weight decay	0.01
Optimiser (beta1, beta2, epsilon)	AdamW (0.9, 0.999, 1e-08)
Monitoring metric (delta, patience)	Validation loss (0.1, 3)

Fine-tuning in a different dataset or style

If you want to fine-tune your own dialogue model, we recommend you to start from the DialoGPT model. You can check the original GitHub repository.

Limitations

This model uses the original English-based tokenizer from GPT-2 paper. Spanish tokenization is not considered but it has similarities in grammatical structure for encoding text. This overlap may help the model transfer its knowledge from English to Spanish. Moreover, the BPE (Byte Pair Encoding) implementation of the GPT-2 tokenizer can assign a representation to every Unicode string. From the GPT-2 paper:

Since our approach can assign a probability to any Unicode string, this allows us to evaluate our LMs on any dataset regardless of pre-processing, tokenization, or vocab size.
This model is intended to be used just for single-turn chitchat conversations in Spanish.
This model's generation capabilities are limited to the extent of the aforementioned fine-tuning dataset.
This model generates short answers, providing general context dialogue in a professional style for the Spanish language.