Bug/Issue: October 2024 movetext starts with "0...", with black's first move starting with "1. "
I'm not sure if this is a problem with the upstream Lichess game database or not, but while working with this data I noticed that the movetext in the October 2024 partition seems to be incorrect. The movetext of every game in that month starts with "0..." for white's first ply, and "1. " marks black's first ply. Haven't checked the entire dataset to see if there are any other month partitions with the problem but I know that at least November 2024 through April 2025 do not have it.
As @nfiekas pointed out, it's likely related to a bug that was fixed in the main db but not in the hf export by this commit: https://github.com/lichess-org/database/commit/bb54a07c3575969a5b37404c681becee834f299e
I'll re-export the data when I get a minute next week.
Thanks again for flagging!
@christopher thank you so much for all the work you do maintaining this data!
I ran a script to scan the rest of the data for any issues like this and found one more month that seems to be affected by the same issue: March 2018
Thank you so much for re-exporting these! Unrelated, but also thank you for changing the shard size to 1 GB. The scripts I run to preprocess and compress the data for my use case run way faster for the 2025 months than they did for comparable size months with smaller shard sizes, it's great!
@christopher is there a way for us community members to help contribute to this dataset? I'd be happy to help keep it up-to-date (bringing in October 2025 through April 2026) as well as check to see if the movetext issues from March 2018 and October 2024 are corrected in the database dumps in https://database.lichess.org/#standard_games and, if so, reupload the fixed versions here, but I'm not entirely sure how to upload to this dataset.
Hi! Sorry this took longer than I anticipated. We've currently paused uploads because we're trying to come up with a better compression and encoding scheme for the data. I'll post an update hopefully in the second end of May.
Sounds great! My offer still stands, if there's anything I can do to help, please let me know! I'm a data scientist but have done some data engineering work, and did a lot of benchmarking work on this data specifically when figuring out the best way to compress it after adding the metadata I needed for my side project (e.g. num_moves, has_eval, avg_elo), unsurprisingly ended up falling back to zstd after flirting w/ brotli for the better ratios but decided the extra time wasn't worth it for me. Only recently found that I had basically just replicated what Niklas had done here: https://backscattering.de/chess/compression/ for this commit to the database repo: https://github.com/lichess-org/database/commit/629c8176bd68e89a4321aadd314bafac2108cca3
Also happy to donate compute time if you have a script that just needs to take days/weeks to run for backfilling periods.