Update README.md

README.md

@@ -15,3 +15,77 @@ configs:
   - split: train
     path: data/train-*
 ---
+
+This dataset was created using the [lichess elite database](https://database.nikonoel.fr/). 
+The it includes the all games up to December 2024.
+
+
+
+After downloading the zip files, the dataset was processed using the following fish script. 
+The goal was to create three UCI-encoded datasets from the lichess elite datset, where all games are deduplicated across all three datasets:
+  1. has_promote: includes non-checkmate games that include pawn promotions
+  1. no_promote: includes non-checkmate games without pawn promotions
+  1. checkmates: includes games that end in checkmate
+```sh
+#!/usr/bin/env fish
+
+# Deduplicate all Lichess elite database files
+pgn-extract --notags --nocomments --nonags --novars -w10000 --noduplicates -o all_deduplicated.san -f lichess_file_list.txt 2>dedupe_output.txt
+
+echo "Deduplication complete. Output saved to all_deduplicated.san."
+
+# Partition games into checkmates and non-checkmates
+pgn-extract --notags --nocomments --nonags --novars -w100000 --checkmate -o checkmates.san -n others.san all_deduplicated.san  2>/dev/null
+
+echo "Games partitioned: checkmates.san and others.san created."
+
+# Further partition non-checkmate games based on pawn promotions
+grep -B1 "=" others.san > has_promote.san
+grep -B1 -v "=" others.san > no_promote.san
+
+echo "Non-checkmate games split into has_promote.san and no_promote.san."
+
+# Convert each SAN file to UCI format
+pgn-extract -Wlalg --noresults --nochecks --nomovenumbers --notags --nocomments --nonags --novars -w100000 -o has_promote.uci has_promote.san 2>/dev/null
+pgn-extract -Wlalg --noresults --nochecks --nomovenumbers --notags --nocomments --nonags --novars -w100000 -o no_promote.uci no_promote.san 2>/dev/null
+pgn-extract -Wlalg --noresults --nochecks --nomovenumbers --notags --nocomments --nonags --novars -w100000 -o checkmates.uci checkmates.san 2>/dev/null
+
+echo "SAN files converted to UCI format."
+
+# Add "#" to the end of each move sequence in checkmates.uc
+sed -i '/./s/$/#/' checkmates.uci
+
+echo "Checkmate sequences updated with '#' as EOS token."
+
+# Remove all blank lines
+sed -i '/^$/d' has_promote.uci
+sed -i '/^$/d' no_promote.uci
+sed -i '/^$/d' checkmates.uci
+
+echo "Blank lines removed. Finished."
+```
+
+Once the raw files were processed, the UCI-encoded files contained the following number of games:
+
+- checkmates.uci:  3,708,644 games that end in checkmate
+- no_promote.uci: 23,201,987 games that do not end in checkmate and have no pawn promotions
+- has_promote.uci:   361,652 games that do not end in checkmate and have at least one pawn promotion
+
+I wanted to balance the number of games from each category with a bias toward games that end in checkmate (so the LLM learns to finish the game rather than simply carry-on playing without a goal).
+To accomplish this, I selected all games from `checkmates.uci` and `has_promote.uci`, and selected games from the `no_promote.uci` at a 5:1 ratio to the number of games in `has_promote.uci`, i.e., 1,808,260 games.
+A simple python script for this is:
+
+```python
+from datasets import load_dataset, concatenate_datasets
+
+checkmates = load_dataset("text", data_files="lichess-elite/checkmates.uci")["train"]
+no_promote = load_dataset("text", data_files="lichess-elite/no_promote.uci")["train"]
+has_promote = load_dataset("text", data_files="lichess-elite/has_promote.uci")["train"]
+
+# shuffle to ensure we're selecting across the entire dataset
+no_promote_subset = no_promote.shuffle(seed=42).select(range(5 * len(has_promote)))
+
+# shuffle entire dataset
+ds = concatenate_datasets([checkmates, has_promote, no_promote_subset]).shuffle(seed=42)
+```
+