You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
In the past, we've successfully translated from the NLLB base model. I believe the only required arguments for translate.py were
src-iso
trg-iso
src-project
clearml-queue and
experiment
There also needed to be a minimal config file in the experiment folder to identify the model to use. When I use this set-up now, I am consistently seeing mixed-language output. It seems like the language codes may not be passed through correctly. Has something changed, or is it possible I've forgotten a crucial step?
A couple of recent examples (see screenshot of mixed-language output below):
"S:\MT\experiments\Indonesia\Behoa\NLLB.3.3B.en-TBBe15.id-TBBe_2025_FEB" with command -m silnlp.nmt.translate --checkpoint base --src-project TBBe15_2024_11_06 --books 1CO 2CO REV --trg-iso ind_Latn --src-iso eng_Latn --clearml-queue production Indonesia\Behoa\NLLB.3.3B.en-TBBe15.id-TBBe_2025_FEB and ClearML process (I have tried this with and without the "--checkpoint" flag)
"S:\MT\experiments\Demo_Bethany\NLLB_base" with command -m silnlp.nmt.translate --checkpoint base --src-project NIV11R --books 1JN --trg-iso npi_Deva --src-iso eng_Latn --clearml-queue production Demo_Bethany\NLLB_base for ClearML process
An older example that worked correctly:
"S:\MT\experiments\FT-MalayCentral\NLLB.3.3B.en_NIrV-ind_Latn" with command -m silnlp.nmt.translate --checkpoint best --src-project NIrV --books LUK --trg-iso ind_Latn --src-iso eng_Latn --clearml-queue jobs_urgent FT-MalayCentral\NLLB.3.3B.en_NIrV-ind_Latn and ClearML process
The text was updated successfully, but these errors were encountered:
@bhartmoore Is this blocking experiments? How important is this issue?
I'd say it's important-not-urgent. No project work is currently blocked. There were a couple teams recently for whom I wanted to generate drafts this way, but they have found workarounds. (E.g., one team wanted to convert an English back translation to Indonesian, but will fine tune and translate vernacular-to-Indonesian instead). We do anticipate needing this in the future for project work.
In the past, we've successfully translated from the NLLB base model. I believe the only required arguments for
translate.py
wereThere also needed to be a minimal config file in the experiment folder to identify the model to use. When I use this set-up now, I am consistently seeing mixed-language output. It seems like the language codes may not be passed through correctly. Has something changed, or is it possible I've forgotten a crucial step?
A couple of recent examples (see screenshot of mixed-language output below):
"S:\MT\experiments\Indonesia\Behoa\NLLB.3.3B.en-TBBe15.id-TBBe_2025_FEB" with command
-m silnlp.nmt.translate --checkpoint base --src-project TBBe15_2024_11_06 --books 1CO 2CO REV --trg-iso ind_Latn --src-iso eng_Latn --clearml-queue production Indonesia\Behoa\NLLB.3.3B.en-TBBe15.id-TBBe_2025_FEB
and ClearML process (I have tried this with and without the "--checkpoint" flag)"S:\MT\experiments\Demo_Bethany\NLLB_base" with command
-m silnlp.nmt.translate --checkpoint base --src-project NIV11R --books 1JN --trg-iso npi_Deva --src-iso eng_Latn --clearml-queue production Demo_Bethany\NLLB_base
for ClearML processAn older example that worked correctly:
"S:\MT\experiments\FT-MalayCentral\NLLB.3.3B.en_NIrV-ind_Latn" with command
-m silnlp.nmt.translate --checkpoint best --src-project NIrV --books LUK --trg-iso ind_Latn --src-iso eng_Latn --clearml-queue jobs_urgent FT-MalayCentral\NLLB.3.3B.en_NIrV-ind_Latn
and ClearML processThe text was updated successfully, but these errors were encountered: