Model Card for Aconoya/Nono_instruct_neo-125m_dpo

Este modelo es parte de nuestro esfuerzo por dar acceso a la IA en Español a todas las personas y no sólo a aquellos con un celular de gama alta o una computadora poderosa.

Creado usando gpt-neo-125m por EleutherAI como modelo base y el conjunto de datos conversacional del proyeto OpenAssistant (oasst1) para adaptarlo al formato de instrucciones.

-This model is part of an effort to make AI accessible in Spanish to everyone, not just those with high-end smartphones or powerful computers.-

Model Details

Model Description

• Developed by: Javier R. García
• Language(s) (NLP): English, Spanish
• License: MIT
• Finetuned from: gpt-neo-125m by EleutherAI

Uses

Text generation.

How to Get Started with the Model

Use the code below to get started with the model.

from transformers import AutoModelForCausalLM, AutoConfig, AutoTokenizer
import torch

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model_name = "Aconoya/Nono_instruct_neo-125m_dpo"
model = AutoModelForCausalLM.from_pretrained(model_name)
model.gradient_checkpointing_enable() 
model = model.to(device.type)
tokenizer = AutoTokenizer.from_pretrained(model_name)
tokenizer.pad_token = tokenizer.eos_token

ender_string = '<endofturn>'
system_string='<system>'
user_string='<user>'
assistant_string='<assistant>'

prompt = ['Hello! How are you?', '¡Hola!, ¿Cómo estás?', '¿Qué es un perro?', 'What is a dog?']
prompt = choice(prompt)
formatted_prompt = system_string + 'You are a digital assistant.' + ender_string + '\n' + user_string + prompt + ender_string + '\n' + assistant_string
model_input = tokenizer.encode(formatted_prompt, return_tensors='pt').to(device)
generated_ids = model.generate(input_ids=model_input, pad_token_id=tokenizer.eos_token_id, max_new_tokens=50)
generated_text = tokenizer.decode(generated_ids[:, model_input.shape[-1]:][0], skip_special_tokens=True)
print('Prompt:', prompt)
print("Response: '{}'".format(generated_text))

Training Details

Training Data

The model was trained using the conversational dataset from the OpenAssistant project (oasst1).

Technical Specifications.

The model was trained using Kaggle.com free services.