---
license: apache-2.0
datasets:
- ruanchaves/faquad-nli
language:
- pt
metrics:
- accuracy
library_name: transformers
pipeline_tag: text-classification
tags:
- textual-entailment
widget:
- text: "<s>Qual a capital do Brasil?<s>A capital do Brasil é Brasília!</s>"
  example_title: Exemplo
- text: "<s>Qual a capital do Brasil?<s>Anões são muito mais legais do que elfos!</s>"
  example_title: Exemplo
---
# TeenyTinyLlama-160m-FaQuAD-NLI

TeenyTinyLlama is a pair of small foundational models trained in Brazilian Portuguese.

This repository contains a version of [TeenyTinyLlama-160m](https://huggingface.co/nicholasKluge/TeenyTinyLlama-160m) (`TeenyTinyLlama-160m-FaQuAD-NLI`) fine-tuned on the [FaQuAD-NLI dataset](https://huggingface.co/datasets/ruanchaves/faquad-nli).

## Details

- **Number of Epochs:** 3
- **Batch size:** 16
- **Optimizer:** `torch.optim.AdamW` (learning_rate = 4e-5, epsilon = 1e-8)
- **GPU:** 1 NVIDIA A100-SXM4-40GB

## Usage

Using `transformers.pipeline`:

```python
from transformers import pipeline

text = "<s>Qual a capital do Brasil?<s>A capital do Brasil é Brasília!</s>"

classifier = pipeline("text-classification", model="nicholasKluge/TeenyTinyLlama-160m-FaQuAD-NLI")
classifier(text)

# >>> [{'label': 'SUITABLE', 'score': 0.9774010181427002}]
```

## Reproducing

To reproduce the fine-tuning process, use the following code snippet:

```python
# Faquad-nli
! pip install transformers datasets evaluate accelerate -q

import evaluate
import numpy as np
from datasets import load_dataset, Dataset, DatasetDict
from transformers import AutoTokenizer, DataCollatorWithPadding
from transformers import AutoModelForSequenceClassification, TrainingArguments, Trainer

# Load the task
dataset = load_dataset("ruanchaves/faquad-nli")

# Create a `ModelForSequenceClassification`
model = AutoModelForSequenceClassification.from_pretrained(
    "nicholasKluge/TeenyTinyLlama-160m", 
    num_labels=2, 
    id2label={0: "UNSUITABLE", 1: "SUITABLE"}, 
    label2id={"UNSUITABLE": 0, "SUITABLE": 1}
)

tokenizer = AutoTokenizer.from_pretrained("nicholasKluge/TeenyTinyLlama-160m")

# Format the dataset
train = dataset['train'].to_pandas()
train['text'] = train['question'] + tokenizer.bos_token + train['answer'] + tokenizer.eos_token
train = train[['text', 'label']]
train.labels = train.label.astype(int)
train = Dataset.from_pandas(train)

test = dataset['test'].to_pandas()
test['text'] = test['question'] + tokenizer.bos_token + test['answer'] + tokenizer.eos_token
test = test[['text', 'label']]
test.labels = test.label.astype(int)
test = Dataset.from_pandas(test)

dataset = DatasetDict({
    "train": train,  
    "test": test                  
})

# Preprocess the dataset
def preprocess_function(examples):
    return tokenizer(examples["text"], truncation=True)

dataset_tokenized = dataset.map(preprocess_function, batched=True)

# Create a simple data collactor
data_collator = DataCollatorWithPadding(tokenizer=tokenizer)

# Use accuracy as evaluation metric
accuracy = evaluate.load("accuracy")

# Function to compute accuracy
def compute_metrics(eval_pred):
    predictions, labels = eval_pred
    predictions = np.argmax(predictions, axis=1)
    return accuracy.compute(predictions=predictions, references=labels)

# Define training arguments
training_args = TrainingArguments(
    output_dir="checkpoints",
    learning_rate=4e-5,
    per_device_train_batch_size=16,
    per_device_eval_batch_size=16,
    num_train_epochs=3,
    weight_decay=0.01,
    evaluation_strategy="epoch",
    save_strategy="epoch",
    load_best_model_at_end=True,
    push_to_hub=True,
    hub_token="your_token_here",
    hub_model_id="username/model-ID"
)

# Define the Trainer
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=dataset_tokenized["train"],
    eval_dataset=dataset_tokenized["test"],
    tokenizer=tokenizer,
    data_collator=data_collator,
    compute_metrics=compute_metrics,
)

# Train!
trainer.train()

```

## Fine-Tuning Comparisons

| Models                                                                                     | [FaQuAD-NLI](https://huggingface.co/datasets/ruanchaves/faquad-nli) |
|--------------------------------------------------------------------------------------------|---------------------------------------------------------------------|
| [Bert-base-portuguese-cased](https://huggingface.co/neuralmind/bert-base-portuguese-cased) | 93.07                                                               |
| [Bert-large-portuguese-cased](https://huggingface.co/neuralmind/bert-base-portuguese-cased)| 92.26                                                               |
| [Teeny Tiny Llama 460m](https://huggingface.co/nicholasKluge/TeenyTinyLlama-460m)          | 91.18                                                               |
| [Teeny Tiny Llama 160m](https://huggingface.co/nicholasKluge/TeenyTinyLlama-160m)          | 90.00                                                               |
| [Gpt2-small-portuguese](https://huggingface.co/pierreguillou/gpt2-small-portuguese)        | 86.46                                                               |

## Cite as 🤗

```latex

@misc{nicholas22llama,
  doi = {10.5281/zenodo.6989727},
  url = {https://huggingface.co/nicholasKluge/TeenyTinyLlama-160m},
  author = {Nicholas Kluge Corrêa},
  title = {TeenyTinyLlama},
  year = {2023},
  publisher = {HuggingFace},
  journal = {HuggingFace repository},
}

```

## Funding

This repository was built as part of the RAIES ([Rede de Inteligência Artificial Ética e Segura](https://www.raies.org/)) initiative, a project supported by FAPERGS - ([Fundação de Amparo à Pesquisa do Estado do Rio Grande do Sul](https://fapergs.rs.gov.br/inicial)), Brazil.

## License

TeenyTinyLlama-160m-FaQuAD-NLI is licensed under the Apache License, Version 2.0. See the [LICENSE](LICENSE) file for more details.