enable multiple eval datasets#1052
Conversation
|
@younesbelkada |
|
I have the same problem. When multiple eval datasets were pass as dict, It cause an error in dataset.map trl/trl/trainer/sft_trainer.py Line 379 in a60ceef I think this PR is really useful. |
Yes, exactly this. And this error occurs naturally both when using packing and when not. |
lvwerra
left a comment
lvwerra
left a comment
There was a problem hiding this comment.
Looks good to me but I'll let @younesbelkada have a look as well!
|
If you could add a test for this that would be awesome! |
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. |
|
@lvwerra |
|
@lvwerra |
|
@lvwerra @younesbelkada |
younesbelkada
left a comment
younesbelkada
left a comment
There was a problem hiding this comment.
Looks great to me, thanks for this contribution!
|
@peter-sk the tests are taking forever, I think that something went wrong with the test you designed, can you please have a quick look? |
|
@younesbelkada |
* enable multiple eval datasets * added test * try to avoid infinite computation * make sure eval set is not infinite * downsizing the test
* enable multiple eval datasets * added test * try to avoid infinite computation * make sure eval set is not infinite * downsizing the test
5 participants
The standard Trainer class from the transformers library (and the documentation of SFTTrainer) allow for multiple validation datasets to be passed as a dictionary from dataset name to dataset.
This does not work however in current SFTTrainer code. This PR fixes this.