Datasets:
Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,26 @@
|
|
1 |
-
---
|
2 |
-
license: unknown
|
3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: unknown
|
3 |
+
task_categories:
|
4 |
+
- translation
|
5 |
+
language:
|
6 |
+
- ja
|
7 |
+
- zh
|
8 |
+
---
|
9 |
+
# Derived from larryvrh/CCMatrix-v1-Ja_Zh-filtered
|
10 |
+
Made some changes to the dataset to train the new mt5-base model
|
11 |
+
* since it's all from the community anyway so i disclose this.
|
12 |
+
* putting lora adapters isn't sufficient to solve old and general habits
|
13 |
+
## weakness of the translation model
|
14 |
+
* translation stops or repeats after 30 words, and doesnt recognize line breaks
|
15 |
+
* the dataset generally too short, 83% below 50 words
|
16 |
+
* solution: fused some sentenses with " ","。" or line breaks to make them longer
|
17 |
+
* now it has similar percentage of each length
|
18 |
+
* mt5 doesn't handle well above 250 due to default length
|
19 |
+
* model can't decide which “ to use, and it randomly add or remove non-words or numbers
|
20 |
+
* the dataset itself dirty with this aspect, becomes old habit
|
21 |
+
* solution: filtered all the data where two side don't match on number of non-words
|
22 |
+
* they are cool as lora features, but I don't want them in the base
|
23 |
+
* model has habits of not translation words when translate item descriptions
|
24 |
+
* there are some description-like samples in the dataset, with untranslated ja characters
|
25 |
+
* solution: removed them
|
26 |
+
* they are mistakes
|