Chemically modified small interfering RNAs (siRNAs) are promising therapeutics guiding sequence-specific silencing of disease genes. However, identifying chemically modified siRNA sequences that effectively silence target genes is a challenge. Such determinations necessitate computational algorithms. Machine Learning (ML) is a powerful predictive approach for tackling biological problems, but typically requires datasets significantly larger than most available siRNA datasets. Here, we describe a framework for applying ML to a small dataset (356 modified sequences) for siRNA efficacy prediction. To overcome noise and biological limitations in siRNA datasets, we apply a trichotomous (using two thresholds) partitioning approach, producing several combinations of classification threshold pairs. We then test the effects of different thresholds on random forest (RF) ML model performance using a novel evaluation metric accounting for class imbalances. We identify thresholds yielding a model with high predictive power outperforming a simple linear classification model generated from the same data. Using a novel method to extract model features, we observe target site base preferences consistent with current understanding of the siRNA-mediated silencing mechanism, with RF providing higher resolution than the linear model. This framework applies to any classification challenge involving small biological datasets, providing an opportunity to develop high-performing design algorithms for oligonucleotide therapies.