Thanks for the suggestion! I just went back and added it to my benchmarking script. If I leave out the stri_replace_last_regex function call, it is pretty similar to vroom (5.7 vs 5.8 ms), which is a smidge faster than dt (7.6 ms). With the stri_replace_last_regex function call it is still similar to vroom (33 ms), but a smidge slower than dt (30.7 ms). I committed the additional test to the repository if you want to see what I did.