Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,38 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language:
|
| 3 |
+
- ru
|
| 4 |
+
- uk
|
| 5 |
+
- be
|
| 6 |
+
- kk
|
| 7 |
+
- az
|
| 8 |
+
- hy
|
| 9 |
+
- ka
|
| 10 |
+
- he
|
| 11 |
+
- en
|
| 12 |
+
- de
|
| 13 |
+
tags:
|
| 14 |
+
- language classification
|
| 15 |
+
datasets:
|
| 16 |
+
- open_subtitles
|
| 17 |
+
- tatoeba
|
| 18 |
+
- oscar
|
| 19 |
+
---
|
| 20 |
+
|
| 21 |
+
# RoBERTa for Multilabel Language Classification
|
| 22 |
+
## Training
|
| 23 |
+
RoBERTa fine-tuned on small parts of Open Subtitles, Oscar and Tatoeba datasets (~9k samples per language).
|
| 24 |
+
|
| 25 |
+
Implemented heuristic algorithm for multilingual training data creation - https://github.com/n1kstep/lang-classifier
|
| 26 |
+
|
| 27 |
+
| data source | language |
|
| 28 |
+
|-----------------|----------------|
|
| 29 |
+
| open_subtitles | ka, he, en, de |
|
| 30 |
+
| oscar | be, kk, az, hu |
|
| 31 |
+
| tatoeba | ru, uk |
|
| 32 |
+
|
| 33 |
+
## Validation
|
| 34 |
+
The metrics obtained from validation on the another part of dataset (~1k samples per language).
|
| 35 |
+
|
| 36 |
+
| Training Loss | Validation Loss | F1-Score | Roc Auc | Accuracy | Support |
|
| 37 |
+
|---------------|-----------------|----------|----------|----------|---------|
|
| 38 |
+
| 0.161500 | 0.110949 | 0.947844 | 0.953939 | 0.762063 | 26858 |
|