I see this question is really old, but just my two cents: You could use the F1 score. I usually use it for imbalanced/skewed datasets, as it represents how well the model performs on both precision and recall - making it specifically meaningful for skewed datasets.
Whether any accuracy measure is going to be meaningful depends on how well the small sample represents the larger dataset. This is hard to know, because your data may contain complexities that cannot be represented in smaller subsamples of the larger dataset.
Obviously, another issue is the sample size itself. 25 samples is a really small sample size for any learning task. I could imagine a learning task on such a small sample size to be successful on the condition that, for instance, your prediction target classes are linearly separable (the data is of low complexity), in which case the learning task is easy, and sample size becomes less important. This is rarely the case in my experience with real world data.
In brief: You could use the F1 score, if the sample is representative (hard to know for sure), but the small sample size will almost certainly make this a very hard problem - unless the data are non-complex (e.g. output classes are linearly separable).