IDEA-Bench-Arena / arena_elo /update_elo_rating.sh
自良
update log server
0ce740f
mkdir -p results
# for battle data
python -m elo_rating.clean_battle_data
battle_cutoff_date=`cat cut_off_date.txt` && rm cut_off_date.txt && echo "Battle data last updated on $battle_cutoff_date"
mkdir -p ./results/$battle_cutoff_date
cp clean_battle_$battle_cutoff_date.json ./results/latest/clean_battle.json
mv clean_battle_$battle_cutoff_date.json ./results/$battle_cutoff_date/clean_battle.json
python -m elo_rating.elo_analysis --clean-battle-file ./results/$battle_cutoff_date/clean_battle.json
mv ./elo_results_$battle_cutoff_date.pkl ./results/$battle_cutoff_date/elo_results.pkl
# generate the leaderboard
python -m elo_rating.generate_leaderboard \
--model_info_file "../model/model_registry.csv" \
--elo_rating_pkl "./results/$battle_cutoff_date/elo_results.pkl" \
--output_csv "./results/$battle_cutoff_date/leaderboard.csv"
mkdir -p ./results/latest
cp ./results/$battle_cutoff_date/leaderboard.csv ./results/latest/leaderboard.csv
cp ./results/$battle_cutoff_date/elo_results.pkl ./results/latest/elo_results.pkl