Is it possible to make the model training log public?

[asirgogogo]

model：https://huggingface.co/Dongwei/Qwen-2.5-7B_Base_Math_smalllr

[Some-random]

I'm not sure how to make my runs/project public, but here is a screenshot, hope it helps!

[asirgogogo]

@Some-random
I observed that your training step is 68. When I was experimenting, I used the same parameter configuration and the training step number was 468

trl 0.15.0.dev0
transformers 4.49.0.dev0

[merlinarer]

@Some-random
Hello, how do you evaluate model in yr setting? I use lighteval in this repo and get a very low acc.

achieves 69.4% accuracy on MATH-500, demonstrating a 17%+ improvement over the base model.

[Some-random]

@asirgogogo I'm only training on MATH. You can check the config I used to train this model

[Some-random]

@merlinarer My lighteval version is lighteval @ git+https://github.com/huggingface/lighteval.git@0e462692436e1f0575bdb4c6ef63453ad9bde7d4, hope that helps. I've also seen many issues around mismatched accuracy numbers... Hope the maintainer will fix it soon