To test scoring functions in protein docking, one can use a statistically significant set of docking poses with at least one near-native match and several false-positive matches (decoys) per complex. Most proteins in large-scale structural modeling of interactome will be themselves models of limited accuracy. So far, the docking decoys sets have been built for X-ray structures only. To our knowledge, this set is the first one generated for systematic arrays of protein models of different structural accuracy. Ideally the decoys should be spread in space to avoid clustering bias. This decoy set was built from 300,000 docking matches generated by GRAMM rigid-body docking scan at all distortion levels - from 0Å (the original crystal structure) to 6Å, for all complexes in the Model Docking Benchmark 2. The set for each complex at each distortion level contains one near-native match, (acceptable of better by CAPRI criteria) top ranked by shape complementarity only, and 99 sparsely distributed incorrect docking matches with energy similar to the near-native pose. All matches are unscored and unrefined. For each complex in this dataset there are two files. The first one is a modeled structure of the receptor, distorted to the level indicated in the file name. The second one is 100 positions of the ligand separated, similar to NMR files, by the MODEL-ENDMDL tags.