Reproducing no-filtering baseline commonpool_s_s13m_b4k

[dwahdany]

I'm trying to replicate basic results without filtering for now. There are some major differences between the models I trained locally and the ones on huggingface. The accuracy of the locally trained models is worse, when doing linear probing the differences are even larger (20% acc. from locally trained model vs. 44% from huggingface model on CIFAR100)

Dataset	Encoder	Zero-shot Test	Linear Probe Test
cifar10	commonpool_s_s13m_b4k	0.4077	0.685 ± 0.0014
cifar10	local_commonpool_s_s13m_b4k_0	0.3572	0.4694 ± 0.0106
cifar10	local_commonpool_s_s13m_b4k_1	0.3443	0.4565 ± 0.0143
cifar10	local_commonpool_s_s13m_b4k_3	0.3406	0.4609 ± 0.0126
cifar10	local_commonpool_s_s13m_b4k_4	0.3346	0.469 ± 0.0141
cifar10	local_commonpool_s_s13m_b4k_2	0.3323	0.4447 ± 0.0164
vtab/cifar100	commonpool_s_s13m_b4k	0.1297	0.4355 ± 0.0025
vtab/cifar100	local_commonpool_s_s13m_b4k_1	0.1246	0.2024 ± 0.0035
vtab/cifar100	local_commonpool_s_s13m_b4k_0	0.1168	0.1997 ± 0.0085
vtab/cifar100	local_commonpool_s_s13m_b4k_3	0.1139	0.2004 ± 0.0066
vtab/cifar100	local_commonpool_s_s13m_b4k_2	0.1138	0.2002 ± 0.0043
vtab/cifar100	local_commonpool_s_s13m_b4k_4	0.1128	0.2047 ± 0.0044

1. To my understanding, just calling train.py --scale small on the unmodified commonpool dataset should replicate the no-filter baseline commonpool_s_s13m_b4k. Is that right?
2. I ran five different seeds for the pretraining and for each ten different seeds for the linear probing. Why are the results so different from the online models?