Commit
•
8c03084
1
Parent(s):
bfde0c2
Update README.md
Browse files
README.md
CHANGED
@@ -16,38 +16,80 @@ metrics:
|
|
16 |
model-index:
|
17 |
- name: dougtrajano/toxicity-type-detection
|
18 |
results: []
|
|
|
|
|
|
|
19 |
---
|
20 |
|
21 |
-
<!-- This model card has been generated automatically according to the information the Trainer had access to. You
|
22 |
-
should probably proofread and complete it, then remove this comment. -->
|
23 |
-
|
24 |
# dougtrajano/toxicity-type-detection
|
25 |
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
-
|
31 |
-
|
32 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
33 |
|
34 |
-
|
35 |
|
36 |
-
|
37 |
|
38 |
-
##
|
39 |
|
40 |
-
|
41 |
|
42 |
-
|
43 |
|
44 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
45 |
|
46 |
## Training procedure
|
47 |
|
48 |
### Training hyperparameters
|
49 |
|
50 |
The following hyperparameters were used during training:
|
|
|
51 |
- learning_rate: 7.044186985160909e-05
|
52 |
- train_batch_size: 8
|
53 |
- eval_batch_size: 8
|
@@ -56,24 +98,13 @@ The following hyperparameters were used during training:
|
|
56 |
- lr_scheduler_type: linear
|
57 |
- num_epochs: 30
|
58 |
|
59 |
-
### Training results
|
60 |
-
|
61 |
-
| Training Loss | Epoch | Step | Validation Loss | Accuracy | F1 | Precision | Recall |
|
62 |
-
|:-------------:|:-----:|:----:|:---------------:|:--------:|:------:|:---------:|:------:|
|
63 |
-
| 1.1107 | 1.0 | 534 | 0.9282 | 0.2823 | 0.6762 | 0.7419 | 0.6630 |
|
64 |
-
| 0.8974 | 2.0 | 1068 | 0.8605 | 0.2754 | 0.6324 | 0.7759 | 0.5913 |
|
65 |
-
| 0.7436 | 3.0 | 1602 | 1.0151 | 0.3150 | 0.6870 | 0.7828 | 0.6512 |
|
66 |
-
| 0.644 | 4.0 | 2136 | 1.1455 | 0.3519 | 0.7114 | 0.7857 | 0.6865 |
|
67 |
-
| 0.4704 | 5.0 | 2670 | 1.4827 | 0.3387 | 0.7109 | 0.7814 | 0.6843 |
|
68 |
-
| 0.3316 | 6.0 | 3204 | 1.6275 | 0.3602 | 0.7217 | 0.8020 | 0.6816 |
|
69 |
-
| 0.2717 | 7.0 | 3738 | 2.2337 | 0.4214 | 0.7645 | 0.8180 | 0.7230 |
|
70 |
-
| 0.231 | 8.0 | 4272 | 2.0275 | 0.3651 | 0.7194 | 0.8271 | 0.6528 |
|
71 |
-
| 0.197 | 9.0 | 4806 | 1.9878 | 0.4033 | 0.7409 | 0.8240 | 0.6812 |
|
72 |
-
|
73 |
-
|
74 |
### Framework versions
|
75 |
|
76 |
- Transformers 4.26.0
|
77 |
- Pytorch 1.10.2+cu113
|
78 |
- Datasets 2.9.0
|
79 |
- Tokenizers 0.13.2
|
|
|
|
|
|
|
|
|
|
16 |
model-index:
|
17 |
- name: dougtrajano/toxicity-type-detection
|
18 |
results: []
|
19 |
+
datasets:
|
20 |
+
- dougtrajano/olid-br
|
21 |
+
library_name: transformers
|
22 |
---
|
23 |
|
|
|
|
|
|
|
24 |
# dougtrajano/toxicity-type-detection
|
25 |
|
26 |
+
Toxicity Type Detection is a model that predicts the type(s) of toxicity(s) in a given text.
|
27 |
+
|
28 |
+
Toxicity Labels: `health`, `ideology`, `insult`, `lgbtqphobia`, `other_lifestyle`, `physical_aspects`, `profanity_obscene`, `racism`, `sexism`, `xenophobia`
|
29 |
+
|
30 |
+
This BERT model is a fine-tuned version of [neuralmind/bert-base-portuguese-cased](https://huggingface.co/neuralmind/bert-base-portuguese-cased) on the [OLID-BR dataset](https://huggingface.co/datasets/dougtrajano/olid-br).
|
31 |
+
|
32 |
+
## Overview
|
33 |
+
|
34 |
+
**Input:** Text in Brazilian Portuguese
|
35 |
+
|
36 |
+
**Output:** Multilabel classification (toxicity types)
|
37 |
+
|
38 |
+
## Usage
|
39 |
+
|
40 |
+
```python
|
41 |
+
from transformers import AutoTokenizer, AutoModelForSequenceClassification
|
42 |
+
|
43 |
+
tokenizer = AutoTokenizer.from_pretrained("dougtrajano/toxicity-type-detection")
|
44 |
+
|
45 |
+
model = AutoModelForSequenceClassification.from_pretrained("dougtrajano/toxicity-type-detection")
|
46 |
+
```
|
47 |
+
|
48 |
+
## Limitations and bias
|
49 |
+
|
50 |
+
The following factors may degrade the model’s performance.
|
51 |
|
52 |
+
**Text Language**: The model was trained on Brazilian Portuguese texts, so it may not work well with Portuguese dialects.
|
53 |
|
54 |
+
**Text Origin**: The model was trained on texts from social media and a few texts from other sources, so it may not work well on other types of texts.
|
55 |
|
56 |
+
## Trade-offs
|
57 |
|
58 |
+
Sometimes models exhibit performance issues under particular circumstances. In this section, we'll discuss situations in which you might discover that the model performs less than optimally, and should plan accordingly.
|
59 |
|
60 |
+
**Text Length**: The model was fine-tuned on texts with a word count between 1 and 178 words (average of 18 words). It may give poor results on texts with a word count outside this range.
|
61 |
|
62 |
+
## Performance
|
63 |
+
|
64 |
+
The model was evaluated on the test set of the [OLID-BR](https://dougtrajano.github.io/olid-br/) dataset.
|
65 |
+
|
66 |
+
**Accuracy:** 0.4214
|
67 |
+
|
68 |
+
**Precision:** 0.8180
|
69 |
+
|
70 |
+
**Recall:** 0.7230
|
71 |
+
|
72 |
+
**F1-Score:** 0.7645
|
73 |
+
|
74 |
+
| Label | Precision | Recall | F1-Score | Support |
|
75 |
+
| :---: | :-------: | :----: | :------: | :-----: |
|
76 |
+
| `health` | 0.3182 | 0.1795 | 0.2295 | 39 |
|
77 |
+
| `ideology` | 0.6820 | 0.6842 | 0.6831 | 304 |
|
78 |
+
| `insult` | 0.9689 | 0.8068 | 0.8805 | 1351 |
|
79 |
+
| `lgbtqphobia` | 0.8182 | 0.5870 | 0.6835 | 92 |
|
80 |
+
| `other_lifestyle` | 0.4242 | 0.4118 | 0.4179 | 34 |
|
81 |
+
| `physical_aspects` | 0.4324 | 0.5783 | 0.4948 | 83 |
|
82 |
+
| `profanity_obscene` | 0.7482 | 0.7509 | 0.7496 | 562 |
|
83 |
+
| `racism` | 0.4737 | 0.3913 | 0.4286 | 23 |
|
84 |
+
| `sexism` | 0.5132 | 0.3391 | 0.4084 | 115 |
|
85 |
+
| `xenophobia` | 0.3333 | 0.4375 | 0.3784 | 32 |
|
86 |
|
87 |
## Training procedure
|
88 |
|
89 |
### Training hyperparameters
|
90 |
|
91 |
The following hyperparameters were used during training:
|
92 |
+
|
93 |
- learning_rate: 7.044186985160909e-05
|
94 |
- train_batch_size: 8
|
95 |
- eval_batch_size: 8
|
|
|
98 |
- lr_scheduler_type: linear
|
99 |
- num_epochs: 30
|
100 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
101 |
### Framework versions
|
102 |
|
103 |
- Transformers 4.26.0
|
104 |
- Pytorch 1.10.2+cu113
|
105 |
- Datasets 2.9.0
|
106 |
- Tokenizers 0.13.2
|
107 |
+
|
108 |
+
## Provide Feedback
|
109 |
+
|
110 |
+
If you have any feedback on this model, please [open an issue](https://github.com/DougTrajano/ToChiquinho/issues/new) on GitHub.
|