forked from Helsinki-NLP/Tatoeba-Challenge
-
Notifications
You must be signed in to change notification settings - Fork 0
/
opus-2021-02-23.yml
39 lines (39 loc) · 996 Bytes
/
opus-2021-02-23.yml
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
release: swe-nor/opus-2021-02-23.zip
release-date: 2021-02-23
dataset-name: opus
modeltype: transformer
pre-processing: normalization + SentencePiece (spm32k,spm32k)
subwords:
- source: spm32k
- target: spm32k
subword-models:
- source: source.spm
- target: target.spm
use-target-labels:
- >>nno<<
- >>nob<<
source-languages:
- swe
target-languages:
- nno
- nob
training-data:
- swe-nno: Tatoeba-train (112253)
- swe-nob: Tatoeba-train (8031330)
validation-data:
- nno-swe: Tatoeba-dev, 8
- nob-swe: Tatoeba-dev, 992
- total size of shuffled dev data: 1000
- devset = top 1000 lines of Tatoeba-dev.src.shuffled!
test-data:
- Tatoeba-test.swe-nor: 562/3682
- Tatoeba-test.swe-nno: 2/10
- Tatoeba-test.swe-nob: 560/3671
BLEU-scores:
- Tatoeba-test.swe-nor: 73.6
- Tatoeba-test.swe-nno: 20.4
- Tatoeba-test.swe-nob: 73.7
chr-F-scores:
- Tatoeba-test.swe-nor: 0.843
- Tatoeba-test.swe-nno: 0.565
- Tatoeba-test.swe-nob: 0.842