Difficulty reproducing reported results: SVM accuracy lower than expected

[jonathan-h1]

Hello,

I attempted to reproduce your pre-training results, but I encountered some discrepancies. Following the dataset construction guidelines and default configurations, I was only able to achieve a classification accuracy of 92.2% with the linear SVM.

I noticed that not all files were included in the train.txt and test.txt files. After adding the missing files, the accuracy improved slightly to 92.4%, but it still falls short of the reported 92.9%.

Could you provide guidance on what I might be missing or doing incorrectly in the dataset construction process or with the default configuration?