Commit
•
de7cac6
1
Parent(s):
71c6ef0
Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,155 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: apache-2.0
|
3 |
+
datasets:
|
4 |
+
- christykoh/imdb_pt
|
5 |
+
language:
|
6 |
+
- pt
|
7 |
+
metrics:
|
8 |
+
- accuracy
|
9 |
+
library_name: transformers
|
10 |
+
pipeline_tag: text-classification
|
11 |
+
tags:
|
12 |
+
- sentiment-analysis
|
13 |
+
widget:
|
14 |
+
- text: "Esqueceram de mim 2 é um dos melhores filmes de natal de todos os tempos."
|
15 |
+
example_title: Exemplo
|
16 |
+
- text: "Esqueceram de mim 2 é o pior filme da franquia inteira."
|
17 |
+
example_title: Exemplo
|
18 |
+
---
|
19 |
+
# TeenyTinyLlama-460m-IMDB
|
20 |
+
|
21 |
+
TeenyTinyLlama is a series of small foundational models trained in Brazilian Portuguese.
|
22 |
+
|
23 |
+
This repository contains a version of [TeenyTinyLlama-460m](https://huggingface.co/nicholasKluge/TeenyTinyLlama-460m) (`TeenyTinyLlama-460m-IMDB`) fine-tuned on the the [IMDB dataset](https://huggingface.co/datasets/christykoh/imdb_pt).
|
24 |
+
|
25 |
+
## Details
|
26 |
+
|
27 |
+
- **Number of Epochs:** 3
|
28 |
+
- **Batch size:** 16
|
29 |
+
- **Optimizer:** `torch.optim.AdamW` (learning_rate = 4e-5, epsilon = 1e-8)
|
30 |
+
- **GPU:** 1 NVIDIA A100-SXM4-40GB
|
31 |
+
|
32 |
+
## Usage
|
33 |
+
|
34 |
+
Using `transformers.pipeline`:
|
35 |
+
|
36 |
+
```python
|
37 |
+
from transformers import pipeline
|
38 |
+
|
39 |
+
text = "Esqueceram de mim 2 é um dos melhores filmes de natal de todos os tempos."
|
40 |
+
|
41 |
+
classifier = pipeline("text-classification", model="nicholasKluge/TeenyTinyLlama-460m-IMDB")
|
42 |
+
classifier(text)
|
43 |
+
|
44 |
+
# >>> [{'label': 'POSITIVE', 'score': 0.9971244931221008}]
|
45 |
+
```
|
46 |
+
|
47 |
+
## Reproducing
|
48 |
+
|
49 |
+
To reproduce the fine-tuning process, use the following code snippet:
|
50 |
+
|
51 |
+
```python
|
52 |
+
# IMDB
|
53 |
+
! pip install transformers datasets evaluate accelerate -q
|
54 |
+
|
55 |
+
import evaluate
|
56 |
+
import numpy as np
|
57 |
+
from datasets import load_dataset
|
58 |
+
from transformers import AutoTokenizer, DataCollatorWithPadding
|
59 |
+
from transformers import AutoModelForSequenceClassification, TrainingArguments, Trainer
|
60 |
+
|
61 |
+
# Load the task
|
62 |
+
dataset = load_dataset("christykoh/imdb_pt")
|
63 |
+
|
64 |
+
# Create a `ModelForSequenceClassification`
|
65 |
+
model = AutoModelForSequenceClassification.from_pretrained(
|
66 |
+
"nicholasKluge/TeenyTinyLlama-460m",
|
67 |
+
num_labels=2,
|
68 |
+
id2label={0: "NEGATIVE", 1: "POSITIVE"},
|
69 |
+
label2id={"NEGATIVE": 0, "POSITIVE": 1}
|
70 |
+
)
|
71 |
+
|
72 |
+
tokenizer = AutoTokenizer.from_pretrained("nicholasKluge/TeenyTinyLlama-460m")
|
73 |
+
|
74 |
+
# Preprocess the dataset
|
75 |
+
def preprocess_function(examples):
|
76 |
+
return tokenizer(examples["text"], truncation=True, max_length=256)
|
77 |
+
|
78 |
+
dataset_tokenized = dataset.map(preprocess_function, batched=True)
|
79 |
+
|
80 |
+
# Create a simple data collactor
|
81 |
+
data_collator = DataCollatorWithPadding(tokenizer=tokenizer)
|
82 |
+
|
83 |
+
# Use accuracy as an evaluation metric
|
84 |
+
accuracy = evaluate.load("accuracy")
|
85 |
+
|
86 |
+
# Function to compute accuracy
|
87 |
+
def compute_metrics(eval_pred):
|
88 |
+
predictions, labels = eval_pred
|
89 |
+
predictions = np.argmax(predictions, axis=1)
|
90 |
+
return accuracy.compute(predictions=predictions, references=labels)
|
91 |
+
|
92 |
+
# Define training arguments
|
93 |
+
training_args = TrainingArguments(
|
94 |
+
output_dir="checkpoints",
|
95 |
+
learning_rate=4e-5,
|
96 |
+
per_device_train_batch_size=16,
|
97 |
+
per_device_eval_batch_size=16,
|
98 |
+
num_train_epochs=3,
|
99 |
+
weight_decay=0.01,
|
100 |
+
evaluation_strategy="epoch",
|
101 |
+
save_strategy="epoch",
|
102 |
+
load_best_model_at_end=True,
|
103 |
+
push_to_hub=True,
|
104 |
+
hub_token="your_token_here",
|
105 |
+
hub_model_id="username/model-name-imdb"
|
106 |
+
)
|
107 |
+
|
108 |
+
# Define the Trainer
|
109 |
+
trainer = Trainer(
|
110 |
+
model=model,
|
111 |
+
args=training_args,
|
112 |
+
train_dataset=dataset_tokenized["train"],
|
113 |
+
eval_dataset=dataset_tokenized["test"],
|
114 |
+
tokenizer=tokenizer,
|
115 |
+
data_collator=data_collator,
|
116 |
+
compute_metrics=compute_metrics,
|
117 |
+
)
|
118 |
+
|
119 |
+
# Train!
|
120 |
+
trainer.train()
|
121 |
+
```
|
122 |
+
|
123 |
+
## Fine-Tuning Comparisons
|
124 |
+
|
125 |
+
| Models | [IMDB](https://huggingface.co/datasets/christykoh/imdb_pt) |
|
126 |
+
|--------------------------------------------------------------------------------------------|------------------------------------------------------------|
|
127 |
+
| [Bert-large-portuguese-cased](https://huggingface.co/neuralmind/bert-base-portuguese-cased)| 93.58 |
|
128 |
+
| [Teeny Tiny Llama 460m](https://huggingface.co/nicholasKluge/TeenyTinyLlama-460m) | 92.28 |
|
129 |
+
| [Bert-base-portuguese-cased](https://huggingface.co/neuralmind/bert-base-portuguese-cased) | 92.22 |
|
130 |
+
| [Gpt2-small-portuguese](https://huggingface.co/pierreguillou/gpt2-small-portuguese) | 91.60 |
|
131 |
+
| [Teeny Tiny Llama 160m](https://huggingface.co/nicholasKluge/TeenyTinyLlama-160m) | 91.14 |
|
132 |
+
|
133 |
+
## Cite as 🤗
|
134 |
+
|
135 |
+
```latex
|
136 |
+
|
137 |
+
@misc{nicholas22llama,
|
138 |
+
doi = {10.5281/zenodo.6989727},
|
139 |
+
url = {https://huggingface.co/nicholasKluge/TeenyTinyLlama-460m},
|
140 |
+
author = {Nicholas Kluge Corrêa},
|
141 |
+
title = {TeenyTinyLlama},
|
142 |
+
year = {2023},
|
143 |
+
publisher = {HuggingFace},
|
144 |
+
journal = {HuggingFace repository},
|
145 |
+
}
|
146 |
+
|
147 |
+
```
|
148 |
+
|
149 |
+
## Funding
|
150 |
+
|
151 |
+
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.
|
152 |
+
|
153 |
+
## License
|
154 |
+
|
155 |
+
TeenyTinyLlama-460m-IMDB is licensed under the Apache License, Version 2.0. See the [LICENSE](LICENSE) file for more details.
|