Coarse Fusion 长跑调参

启动一轮长跑

./scripts/evaluation/start_coarse_fusion_tuning_long.sh

可用环境变量：

MAX_EVALS=48 BATCH_SIZE=3 CANDIDATE_POOL_SIZE=512 \
RUN_NAME=coarse_fusion_long_001 \
./scripts/evaluation/start_coarse_fusion_tuning_long.sh

启动后会打印：

• run_name
• pid
• log
• run_dir

默认搜索空间：

• scripts/evaluation/tuning/coarse_rank_fusion_space.yaml

默认 baseline seed：

• artifacts/search_evaluation/batch_reports/batch_20260415T150754Z_00b6a8aa3d.md

查看进度

tail -f artifacts/search_evaluation/tuning_launches/<run_name>.log
cat artifacts/search_evaluation/tuning_runs/<run_name>/leaderboard.csv
sed -n '1,200p' artifacts/search_evaluation/tuning_runs/<run_name>/summary.md

实时记录文件：

• trials.jsonl
• leaderboard.csv
• summary.json
• summary.md

续跑

./scripts/evaluation/resume_coarse_fusion_tuning_long.sh <run_name>

也可直接传完整目录：

./scripts/evaluation/resume_coarse_fusion_tuning_long.sh \
  artifacts/search_evaluation/tuning_runs/<run_name>

停止

kill "$(cat artifacts/search_evaluation/tuning_launches/<run_name>.pid)"

说明

• 每轮会自动写入 config/config.yaml
• 每轮会自动执行 ./restart.sh backend
• 如果 eval-web 因 backend 重启不可用，调参器会尝试补拉起 eval-web
• 默认不 apply-best，跑完后会恢复 baseline 配置