This study evaluated machine-learning methods for predicting the mechanical properties of Qom Formation conglomerates in the Tajarak area, northeast of Hamedan, Iran. Heterogeneity in clast size and composition, matrix content, cement, and voids complicate representative coring and direct strength determination. Eight large blocks were collected, and their physical properties, P-wave velocity, point-load index, uniaxial compressive strength (UCS), Brazilian tensile strength (BTS), and static elastic modulus (E_s) were determined under dry and saturated conditions. The dataset contained 16 records representing eight blocks under two moisture conditions. OLS, Ridge, SVR, and RF were compared using three predictor sets. To prevent optimistic accuracy estimates, all dry and saturated records of one block were excluded from training in each validation iteration. Uncertainty was assessed using 10,000 complete-block bootstrap repetitions. Predictor importance was evaluated by removing each variable separately and recalculating model error. RF using physical properties provided the best UCS result (R^2=0.50 and MAE=5.12 MPa). RF using physical properties, P-wave velocity, and the size-corrected point-load index I_(s(50)) produced the best BTS prediction (R^2=0.87 and MAE=0.36 MPa). In contrast, Ridge prediction of E_s was weak and uncertain (R^2=0.29). Removing moisture condition caused the largest increase in UCS prediction error, indicating that the dry-saturated difference supplied the most useful information for the UCS model. For BTS, removing either moisture condition or I_(s(50)) increased error more than removing the other predictors. Because the dataset comprised only eight independent blocks, these models are suitable only for preliminary estimation within the study area; direct testing remains necessary for engineering design.