thomasdehaene
commited on
Commit
•
edf505d
1
Parent(s):
ceaefd1
Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,58 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
language:
|
3 |
+
- nl
|
4 |
+
tags:
|
5 |
+
- text-classification
|
6 |
+
- pytorch
|
7 |
+
widget:
|
8 |
+
- text: "Ik heb je lief met heel mijn hart"
|
9 |
+
example_title: "Non toxic comment 1"
|
10 |
+
- text: "Dat is een goed punt, zo had ik het nog niet bekeken."
|
11 |
+
example_title: "Non toxic comment 2"
|
12 |
+
- text: "Wat de fuck zei je net tegen me, klootzak?"
|
13 |
+
example_title: "Toxic comment 1"
|
14 |
+
- text: "Rot op, vuile hoerenzoon."
|
15 |
+
example_title: "Toxic comment 2"
|
16 |
+
license: apache-2.0
|
17 |
+
metrics:
|
18 |
+
- Accuracy, F1 Score, Recall, Precision
|
19 |
+
---
|
20 |
+
# distilbert-base-dutch-toxic-comments
|
21 |
+
|
22 |
+
## Model description:
|
23 |
+
This model was created with the purpose to detect toxic or potentially harmful comments.
|
24 |
+
|
25 |
+
For this model, we finetuned a multilingual distilbert model [distilbert-base-multilingual-cased](https://huggingface.co/distilbert-base-multilingual-cased) on the translated [Jigsaw Toxicity dataset](https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge).
|
26 |
+
|
27 |
+
The original dataset was translated using the appropriate [MariantMT model](https://huggingface.co/Helsinki-NLP/opus-mt-en-nl).
|
28 |
+
|
29 |
+
The model was trained for 2 epochs, on 90% of the dataset, with the following arguments:
|
30 |
+
```
|
31 |
+
training_args = TrainingArguments(
|
32 |
+
learning_rate=3e-5,
|
33 |
+
per_device_train_batch_size=16,
|
34 |
+
per_device_eval_batch_size=16,
|
35 |
+
gradient_accumulation_steps=4,
|
36 |
+
load_best_model_at_end=True,
|
37 |
+
metric_for_best_model="recall",
|
38 |
+
epochs=2,
|
39 |
+
evaluation_strategy="steps",
|
40 |
+
save_strategy="steps",
|
41 |
+
save_total_limit=10,
|
42 |
+
logging_steps=100,
|
43 |
+
eval_steps=250,
|
44 |
+
save_steps=250,
|
45 |
+
weight_decay=0.001,
|
46 |
+
report_to="wandb")
|
47 |
+
```
|
48 |
+
|
49 |
+
## Model Performance:
|
50 |
+
|
51 |
+
Model evaluation was done on 1/10th of the dataset, which served as the test dataset.
|
52 |
+
|
53 |
+
| Accuracy | F1 Score | Recall | Precision |
|
54 |
+
| --- | --- | --- | --- |
|
55 |
+
| 95.75 | 78.88 | 77.23 | 80.61 |
|
56 |
+
|
57 |
+
## Dataset:
|
58 |
+
Unfortunately we cannot open-source the dataset, since we are bound by the underlying Jigsaw license.
|