File size: 983 Bytes
475e27c
d69129b
475e27c
 
 
 
 
 
 
 
 
 
d69129b
475e27c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
---
language:
- ru
- uk
- be
- kk
- az
- hy
- ka
- he
- en
- de
- multilingual
tags:
- language classification
datasets:
- open_subtitles
- tatoeba
- oscar
---

# RoBERTa for Multilabel Language Classification
## Training
RoBERTa fine-tuned on small parts of Open Subtitles, Oscar and Tatoeba datasets (~9k samples per language).

Implemented heuristic algorithm for multilingual training data creation - https://github.com/n1kstep/lang-classifier

| data source          | language          |
|-----------------|----------------|
| open_subtitles    | ka, he, en, de |
| oscar      | be, kk, az, hu |
| tatoeba     | ru, uk         |

## Validation
The metrics obtained from validation on the another part of dataset (~1k samples per language).

| Training Loss | Validation Loss | F1-Score | Roc Auc  | Accuracy | Support |
|---------------|-----------------|----------|----------|----------|---------|
| 0.161500      | 0.110949        | 0.947844 | 0.953939 | 0.762063 | 26858   |