ShaomuTan commited on
Commit
5eb49de
·
1 Parent(s): ad6c753

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +12 -11
README.md CHANGED
@@ -1,10 +1,17 @@
1
- # EC40 MT Dataset -Shaomu Tan
 
 
2
 
3
  ### EC40 is an English-Centric Multilingual Machine Translation Dataset. It has over 60 Million sentences including 40 Languages across 5 Language Families.
4
 
5
- ## DO NOT Change ANYTHING in this folder.
 
 
 
 
 
 
6
 
7
- <br>
8
 
9
  -----
10
  ## Languages and Family
@@ -15,10 +22,6 @@
15
  | Romance | French, Spanish, Italian, Portuguese, Romanian, Occitan, Asturian, Catalan |
16
  | Slavic | Russian, Czech, Polish, Bulgarian, Ukrainian, Serbian, Belarusian, Bosnian |
17
  | Indo-Aryan | Hindi, Bengali, Kannada, Marathi, Sindhi, Gujarati, Nepali, Urdu |
18
- |
19
-
20
- <br>
21
-
22
  -----
23
 
24
  ## Dataset Stats
@@ -30,10 +33,8 @@
30
  | Low | af, lb, ro, oc, uk, sr, sd, gu, ti, am | 100k |
31
  | Extremely-Low | no, is, ast, ca, be, bs, ne, ur, kab, so | 50k |
32
 
33
- <br>
34
-
35
  -----
36
- ## Build Fairseq dataset (Sharded)
37
 
38
  ```
39
  Read toolkit/build_fairseq_sharded_dataset.sh
@@ -46,4 +47,4 @@ Read toolkit/build_fairseq_sharded_dataset.sh
46
 
47
  ```
48
  Read toolkit/train-EC40-mTrans-large.sh
49
- ```
 
1
+ # EC40 MNMT Dataset
2
+
3
+ GitHub: https://github.com/Smu-Tan/ZS-NMT-Variations/tree/main
4
 
5
  ### EC40 is an English-Centric Multilingual Machine Translation Dataset. It has over 60 Million sentences including 40 Languages across 5 Language Families.
6
 
7
+ #### Note: The dataset is cleaned and pre-processed using tools like Moses, for more details, please refer to the paper.
8
+
9
+ ### Features:
10
+ 1. We carefully balanced the dataset across resources and languages by strictly maintaining each resource group containing 5 language families and each family consists of 8 representative languages.
11
+ 2. EC40 covers a wide spectrum of resource availability, ranging from High(5M) to Medium(1M), Low(100K), and extremely-Low(50K) resources.
12
+ 3. In total, there are 80 English-centric directions for training and 1,640 directions (including all supervised and ZS directions) for evaluation.
13
+ 4. We make use of Ntrex-128 and Flores-200 as our validation and test set.
14
 
 
15
 
16
  -----
17
  ## Languages and Family
 
22
  | Romance | French, Spanish, Italian, Portuguese, Romanian, Occitan, Asturian, Catalan |
23
  | Slavic | Russian, Czech, Polish, Bulgarian, Ukrainian, Serbian, Belarusian, Bosnian |
24
  | Indo-Aryan | Hindi, Bengali, Kannada, Marathi, Sindhi, Gujarati, Nepali, Urdu |
 
 
 
 
25
  -----
26
 
27
  ## Dataset Stats
 
33
  | Low | af, lb, ro, oc, uk, sr, sd, gu, ti, am | 100k |
34
  | Extremely-Low | no, is, ast, ca, be, bs, ne, ur, kab, so | 50k |
35
 
 
 
36
  -----
37
+ ## Build Fairseq dataset (Shard->to avoid RAM OOM)
38
 
39
  ```
40
  Read toolkit/build_fairseq_sharded_dataset.sh
 
47
 
48
  ```
49
  Read toolkit/train-EC40-mTrans-large.sh
50
+ ```