-
Notifications
You must be signed in to change notification settings - Fork 2.7k
Open
Description
I have been experimenting with lm_eval
using NeMo models. Based on my understanding, there are two different NeMo formats:
-
Version 1.0
-
Models are saved in the format:
Meta-Llama-3-8B-Instruct.nemo
-
Running
lm_eval
with this format works fine and I don’t face any issues.
-
-
Version 2.0
- The saved model format looks different (see screenshots below).
- When I try to run
lm_eval
with this format, I get errors saying that the config file is missing. i did look inside the folders i could find model.yaml but not model_config.yaml . - Note : This models works fine with regular nemo inference code.
Screenshots:
*** CLI Code :**

nemo 2.0 model format :

error with lm_eval wrt nemo 2.0 model

Question
- Is
lm_eval
currently supported for NeMo version 2.0 models? - If yes, how can I properly configure or resolve the missing config error [manually change the name]?
- If not, is there a roadmap or timeline for adding NeMo v2.0 support to
lm_eval
?
Metadata
Metadata
Assignees
Labels
No labels