Created by: stephenroller
Patch Description There was a performance regression in the JSONL dataset indexer, where when building the index it parsed the jsonl files when it really just needed start-of-line numbers. This was originally done to skip over bad parses.
This PR returns us to the much older behavior of throwing a hard crash, midway during training, when we come to a bad data point. On the positive side, we still report where the file was bad.
We may need to build a separate validator to deal with bad data parses, but at the end of the day, it doesn't feel like "skipping bad data" is the right thing to do.
Testing steps In progress