4 ms·

So ~18.6% trained relative to LLaMa. That's not _nothing_ but it's also not great. Thanks for digging into this!