-
Notifications
You must be signed in to change notification settings - Fork 35
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
* Add CJK one stage config * Add comment
- Loading branch information
Showing
2 changed files
with
41 additions
and
0 deletions.
There are no files selected for viewing
40 changes: 40 additions & 0 deletions
40
pipeline/train/configs/opustrainer/teacher.one-stage.cjk.yml
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,40 @@ | ||
# This config includes one-stage training and CJK specific modifiers | ||
# It's helpful when en -> CJK models stop training too early | ||
|
||
datasets: | ||
original: {dataset0} # Original parallel corpus | ||
backtranslated: {dataset1} # Back-translated data | ||
|
||
stages: | ||
- train | ||
|
||
# Train on a mix until early stopping | ||
# (useful for clean back-translated data produced by a strong model) | ||
train: | ||
- original 0.7 | ||
- backtranslated 0.3 | ||
- until original inf | ||
|
||
# The default values of the modifiers are taken from the paper https://arxiv.org/pdf/2311.14838.pdf | ||
# Please refer to docs/opus-trainer.md for further details | ||
modifiers: | ||
## Insert new sentences composed form Unicode noise | ||
- Noise: 0.0005 | ||
min_word_length: 2 # Minimum word length for each word in the noisy sentence | ||
max_word_length: 5 # Maximum word length for each word in the noisy sentence | ||
max_words: 6 # Maximum number of words in each noisy sentence | ||
# generates inline noise (emojis etc.) matching positions in source and target sentences using alignments | ||
# no spm_vocab argument -> alignments will be removed from Marian input | ||
# we don't use alignments for teacher training | ||
# Tags modifier has to be the last one to remove the alignments | ||
- Tags: 0.005 | ||
custom_detok_src: "icu:{src}" | ||
custom_detok_trg: "icu:{trg}" | ||
augment: 1 | ||
tag: 0 | ||
|
||
|
||
# random seed should be different for different teacher models | ||
seed: {seed} | ||
# parallel sentences + token alignments | ||
num_fields: 3 |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters