File size: 5,694 Bytes
f62cb7b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
---
license: apache-2.0
datasets:
- ruanchaves/hatebr
language:
- pt
metrics:
- accuracy
library_name: transformers
pipeline_tag: text-classification
tags:
- hate-speech
widget:
- text: "Não concordo com a sua opinião."
  example_title: Exemplo
- text: "Pega a sua opinião e vai a merda com ela!"
  example_title: Exemplo
---
# TeenyTinyLlama-460m-HateBR

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

This repository contains a version of [TeenyTinyLlama-460m](https://huggingface.co/nicholasKluge/TeenyTinyLlama-160m) (`TeenyTinyLlama-460m-HateBR`) fine-tuned on the [HateBR dataset](https://huggingface.co/datasets/ruanchaves/hatebr).

## 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 = "Pega a sua opinião e vai a merda com ela!"

classifier = pipeline("text-classification", model="nicholasKluge/TeenyTinyLlama-460m-HateBR")
classifier(text)

# >>> [{'label': 'TOXIC', 'score': 0.9998729228973389}]
```

## Reproducing

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

```python

# Hatebr
! pip install transformers datasets evaluate accelerate -q

import evaluate
import numpy as np
from huggingface_hub import login
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/hatebr")

# Format the dataset
train = dataset['train'].to_pandas()
train = train[['instagram_comments', 'offensive_language']]
train.columns = ['text', 'labels']
train.labels = train.labels.astype(int)
train = Dataset.from_pandas(train)

test = dataset['test'].to_pandas()
test = test[['instagram_comments', 'offensive_language']]
test.columns = ['text', 'labels']
test.labels = test.labels.astype(int)
test = Dataset.from_pandas(test)

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

# Create a `ModelForSequenceClassification`
model = AutoModelForSequenceClassification.from_pretrained(
    "nicholasKluge/TeenyTinyLlama-460m", 
    num_labels=2, 
    id2label={0: "NONTOXIC", 1: "TOXIC"}, 
    label2id={"NONTOXIC": 0, "TOXIC": 1}
)

tokenizer = AutoTokenizer.from_pretrained("nicholasKluge/TeenyTinyLlama-460m")

# 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                                                                                     | [HateBr](https://huggingface.co/datasets/ruanchaves/hatebr) | 
|--------------------------------------------------------------------------------------------|-------------------------------------------------------------|
| [Teeny Tiny Llama 460m](https://huggingface.co/nicholasKluge/TeenyTinyLlama-460m)          | 91.64                                                       | 
| [Bert-large-portuguese-cased](https://huggingface.co/neuralmind/bert-base-portuguese-cased)| 91.57                                                       |
| [Bert-base-portuguese-cased](https://huggingface.co/neuralmind/bert-base-portuguese-cased) | 91.28                                                       |
| [Teeny Tiny Llama 160m](https://huggingface.co/nicholasKluge/TeenyTinyLlama-160m)          | 90.71                                                       | 
| [Gpt2-small-portuguese](https://huggingface.co/pierreguillou/gpt2-small-portuguese)        | 87.42                                                       |

## Cite as 🤗

```latex

@misc{nicholas22llama,
  doi = {10.5281/zenodo.6989727},
  url = {https://huggingface.co/nicholasKluge/TeenyTinyLlama-460m},
  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-460m-HateBR is licensed under the Apache License, Version 2.0. See the [LICENSE](LICENSE) file for more details.