Four datasets for multiclass classification can be downloaded here. The LETTER and USPS datasets come from the UCI repository¹⁾. The MNIST dataset²⁾ is an handwritten digit recognition benchmark. The INEX³⁾ dataset contains scientific articles from 18 journals and proceedings of the IEEE coded in a flat TF/IDF feature space.

The file are compressed as tar.gz archive. Their format is the so called LibSVM/SVM^light/SVM^struct format. Each example is represented by a line in the following format:

<line>    = <target> <feature>:<value> ... <feature>:<value> 
<target>  = <int>
<feature> = <integer> 
<value>   = <float>

The target value and each of the feature/value pairs are separated by a space character. The target value denotes the class of the example. For example, the line 5 1:0.43 3:0.12 9284:0.2 specifies an example belonging to class 5 for which feature number 1 has the value 0.43, feature number 3 has the value 0.12, feature number 9284 has the value 0.2, and all the other features have value 0.