Aira-2-124M

Aira-2 is the second version of the Aira instruction-tuned series. Aira-2-124M is an instruction-tuned GPT-style model based on GPT-2. The model was trained with a dataset composed of prompts and completions generated synthetically by prompting already-tuned models (ChatGPT, Llama, Open-Assistant, etc).

Check our gradio-demo in Spaces.

Details

Size: 124,441,344 parameters
Dataset: Instruct-Aira Dataset
Language: English
Number of Epochs: 5
Batch size: 32
Optimizer: torch.optim.AdamW (warmup_steps = 1e2, learning_rate = 5e-4, epsilon = 1e-8)
GPU: 1 NVIDIA A100-SXM4-40GB
Emissions: 0.25 KgCO2 (Singapore)
Total Energy Consumption: 0.52 kWh

This repository has the notebook used to train this model.

Usage

Three special tokens are used to mark the user side of the interaction and the model's response:

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

tokenizer = AutoTokenizer.from_pretrained('nicholasKluge/Aira-2-124M')
aira = AutoModelForCausalLM.from_pretrained('nicholasKluge/Aira-2-124M')

aira.eval()
aira.to(device)

question =  input("Enter your question: ")

inputs = tokenizer(tokenizer.bos_token + question + tokenizer.sep_token, return_tensors="pt").to(device)

responses = aira.generate(**inputs,
    bos_token_id=tokenizer.bos_token_id,
    pad_token_id=tokenizer.pad_token_id,
    eos_token_id=tokenizer.eos_token_id,
    do_sample=True,
    top_k=50,
    max_length=500,
    top_p=0.95,
    temperature=0.7,
    num_return_sequences=2)

print(f"Question: 👤 {question}\n")

for i, response in  enumerate(responses):
    print(f'Response {i+1}: 🤖 {tokenizer.decode(response, skip_special_tokens=True).replace(question, "")}')

The model will output something like:

>>>Question: 👤 What is the capital of Brazil?

>>>Response 1: 🤖 The capital of Brazil is Brasília.
>>>Response 2: 🤖 The capital of Brazil is Brasília.

Limitations

🤥 Generative models can perpetuate the generation of pseudo-informative content, that is, false information that may appear truthful.

🤬 In certain types of tasks, generative models can produce harmful and discriminatory content inspired by historical stereotypes.

Evaluation

Model (GPT-2)	Average	ARC	TruthfulQA	ToxiGen
Aira-2-124M	38.07	24.57	41.02	48.62
GPT-2	35.37	21.84	40.67	43.62
Aira-2-355M	39.68	27.56	38.53	53.19
GPT-2-medium	36.43	27.05	40.76	41.49
Aira-2-774M	42.26	28.75	41.33	56.70
GPT-2-large	35.16	25.94	38.71	40.85
Aira-2-1B5	42.22	28.92	41.16	56.60
GPT-2-xl	36.84	30.29	38.54	41.70

Evaluations were performed using the Language Model Evaluation Harness (by EleutherAI). The notebook used to make these evaluations is available in the this repo.

Cite as 🤗


@misc{nicholas22aira,
  doi = {10.5281/zenodo.6989727},
  url = {https://huggingface.co/nicholasKluge/Aira-2-124M},
  author = {Nicholas Kluge Corrêa},
  title = {Aira},
  year = {2023},
  publisher = {HuggingFace},
  journal = {HuggingFace repository},
}

License

The Aira-2-124M is licensed under the Apache License, Version 2.0. See the LICENSE file for more details.

nicholasKluge
/

Aira-2-124M