RookWorld/run.sh at main · jorahn/RookWorld · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
cd llm.c/dev/data/rook

echo "Generating data from Lichess human games..."
python generate_lichess.py -p $STOCKFISH_PATH
echo "Generating data from Stockfish selfplay..."
python generate_selfplay.py -p $STOCKFISH_PATH
cd ..

echo "Generating llm.c dataset..."
python rook.py
cd ../..

echo "Training GPT-2 model..."
bash scripts/run_gpt2_124M_rook.sh

cd dev/eval

echo "Running evaluations (BIGbench Checkmate in One, Validation Accuracy, vs Stockfish level 0, Self-play)..."
bash rook_eval_parallel.sh ../../log_gpt2_124M_rook/model_00005000.bin ../data/rook/rook_val_500.txt.bak $STOCKFISH_PATH

#python export_hf.py --input ../../log_gpt2_124M_rook/model_00005000.bin --output ../../log_gpt2_124M_rook/rook_gpt2_124M_hf

# run Google Big Bench Checkmate in One Benchmark
#python rook_bb-cio.py -g --model_path ../../log_gpt2_124M_rook/rook_gpt2_124M_hf/

# evaluate the model on the validation set (legal move %, best move accuracy, top 5 move accuracy)
#python rook_accuracy.py -g --model_path ../../log_gpt2_124M_rook/rook_gpt2_124M_hf/ -d ../data/rook/rook_val_500.txt.bak

# play 50 games with greedy decoding vs stockfish at strength level 0 and 50ms time limit per move
#python rook_vs_stockfish.py -n 50 -g --model_path ../../log_gpt2_124M_rook/rook_gpt2_124M_hf/ -p $STOCKFISH_PATH -s 0 -l 0.05

# play 50 games against itself with sampling (topk=5, temp=0.6) until illegal move or win/loss/draw outcome
#python rook_selfplay.py --model_path ../../log_gpt2_124M_rook/rook_gpt2_124M_hf/