Commit
•
95a5efe
1
Parent(s):
bee5fba
finalize upload
Browse files
README.md
CHANGED
|
@@ -3,24 +3,87 @@ license: cc-by-nc-3.0
|
|
| 3 |
datasets:
|
| 4 |
- FredZhang7/toxi-text-3M
|
| 5 |
pipeline_tag: text-classification
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 6 |
---
|
| 7 |
|
| 8 |
**I have decided to release all auto-moderation models at once sometime in July. The curated datasets for training these models will be avaliable first.**
|
| 9 |
|
| 10 |
<br>
|
| 11 |
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 19 |
|
| 20 |
<br>
|
| 21 |
|
| 22 |
<br>
|
| 23 |
|
| 24 |
-
Models tested: roberta, xlm-roberta, bert-
|
| 25 |
-
|
| 26 |
-
Model chosen based on cost-efficiency and performance: bert-multilingual-cased
|
|
|
|
| 3 |
datasets:
|
| 4 |
- FredZhang7/toxi-text-3M
|
| 5 |
pipeline_tag: text-classification
|
| 6 |
+
language:
|
| 7 |
+
- ar
|
| 8 |
+
- es
|
| 9 |
+
- pa
|
| 10 |
+
- th
|
| 11 |
+
- et
|
| 12 |
+
- fr
|
| 13 |
+
- fi
|
| 14 |
+
- no
|
| 15 |
+
- hu
|
| 16 |
+
- lt
|
| 17 |
+
- ur
|
| 18 |
+
- so
|
| 19 |
+
- pl
|
| 20 |
+
- el
|
| 21 |
+
- mr
|
| 22 |
+
- sk
|
| 23 |
+
- gu
|
| 24 |
+
- he
|
| 25 |
+
- af
|
| 26 |
+
- te
|
| 27 |
+
- ro
|
| 28 |
+
- lv
|
| 29 |
+
- sv
|
| 30 |
+
- ne
|
| 31 |
+
- kn
|
| 32 |
+
- it
|
| 33 |
+
- mk
|
| 34 |
+
- cs
|
| 35 |
+
- en
|
| 36 |
+
- de
|
| 37 |
+
- da
|
| 38 |
+
- ta
|
| 39 |
+
- bn
|
| 40 |
+
- pt
|
| 41 |
+
- sq
|
| 42 |
+
- tl
|
| 43 |
+
- uk
|
| 44 |
+
- bg
|
| 45 |
+
- ca
|
| 46 |
+
- sw
|
| 47 |
+
- hi
|
| 48 |
+
- zh
|
| 49 |
+
- ja
|
| 50 |
+
- hr
|
| 51 |
+
- ru
|
| 52 |
+
- vi
|
| 53 |
+
- id
|
| 54 |
+
- sl
|
| 55 |
+
- cy
|
| 56 |
+
- ko
|
| 57 |
+
- nl
|
| 58 |
+
- ml
|
| 59 |
+
- tr
|
| 60 |
+
- fa
|
| 61 |
+
|
| 62 |
+
tags:
|
| 63 |
+
- nlp
|
| 64 |
---
|
| 65 |
|
| 66 |
**I have decided to release all auto-moderation models at once sometime in July. The curated datasets for training these models will be avaliable first.**
|
| 67 |
|
| 68 |
<br>
|
| 69 |
|
| 70 |
+
| | v2 | v1 |
|
| 71 |
+
|----------|----------|----------|
|
| 72 |
+
| Base Model | bert-base-multilingual-cased | nlpaueb/legal-bert-small-uncased |
|
| 73 |
+
| Base Tokenizer | bert-base-multilingual-cased | bert-base-multilingual-cased |
|
| 74 |
+
| Framework | PyTorch | TensorFlow |
|
| 75 |
+
| Dataset Size | 2.95M | 2.68M |
|
| 76 |
+
| Train Split | 80% English<br>20% English + 100% Multilingual | None |
|
| 77 |
+
| English Train Accuracy | 99.4% | N/A (≈98%) |
|
| 78 |
+
| Final Train Accuracy | 96.5% | 96.6% |
|
| 79 |
+
| Final Val Accuracy | 95.0% | 94.6% |
|
| 80 |
+
| Languages | 55 | N/A (≈35) |
|
| 81 |
+
| Hyperparameters | maxlen=208<br>batch_size=112<br>optimizer=Adam<br>learning_rate=1e-5<br>loss=BCEWithLogitsLoss() | maxlen=192<br>batch_size=16<br>optimizer=Adam<br>learning_rate=1e-5<br>loss="binary_crossentropy" |
|
| 82 |
+
| Training Stopped | 6/30/2023 | 9/05/2022 |
|
| 83 |
|
| 84 |
<br>
|
| 85 |
|
| 86 |
<br>
|
| 87 |
|
| 88 |
+
Models tested for v2: roberta, xlm-roberta, bert-small, bert-base-cased/uncased, bert-multilingual-cased/uncased, and alberta-large-v2.
|
| 89 |
+
From these models, I chose bert-multilingual-cased because of its higher resource efficiency and performance than the rest for this particular task.
|
|
|