chore: add FORCE flag to split script; align train config pool sizes to 0.4 split

This commit is contained in:
2026-07-16 07:46:31 -04:00
parent 5fdcd9b5aa
commit 933de6251e
2 changed files with 11 additions and 3 deletions
+3 -2
View File
@@ -40,10 +40,11 @@ harness:
skill_update_mode: patch skill_update_mode: patch
appendix_consolidate_threshold: 6 appendix_consolidate_threshold: 6
# 池配置 —— global 冻结切分(diag/val/test 尺寸由 pools.json 冻结,以下采样旋钮加载时忽略) # 池配置 —— global 冻结切分(diag/val/test 尺寸由 pools.json 冻结,以下采样旋钮加载时忽略)
# 实际冻结切分(0.4+tierfingerprint d456ef4f: diag=180 val=120 test=600
pool_split_mode: global pool_split_mode: global
diag_size: 210 diag_size: 180
diag_correct_ratio: 0.5 diag_correct_ratio: 0.5
val_size: 90 val_size: 120
val_correct_ratio: 0.5 val_correct_ratio: 0.5
test_size: 600 test_size: 600
test_questions: "benchmarks/Video-MME" test_questions: "benchmarks/Video-MME"
+8 -1
View File
@@ -50,6 +50,13 @@ if [ "${MODE:-}" = "mock" ]; then
DRY_RUN_FLAG="--dry-run" DRY_RUN_FLAG="--dry-run"
fi fi
# FORCE=1 → --force:已存在冻结产物时覆盖(旧产物自动备份为 .bak.*,WP2 覆盖保护)。
# 重跑切分(改 val_ratio / tier 感知 / 换指纹)时需要;缺省拒绝覆盖以防误毁冻结锚点。
FORCE_FLAG=""
if [ "${FORCE:-}" = "1" ]; then
FORCE_FLAG="--force"
fi
# 诊断并发上限(asyncio.Semaphore):默认 12,可用 CONCURRENCY 环境变量覆盖。 # 诊断并发上限(asyncio.Semaphore):默认 12,可用 CONCURRENCY 环境变量覆盖。
# 注:LLM 熔断阈值/冷却是工程配置,走 .envLLM_CIRCUIT_BREAKER_THRESHOLD / _COOLDOWN), # 注:LLM 熔断阈值/冷却是工程配置,走 .envLLM_CIRCUIT_BREAKER_THRESHOLD / _COOLDOWN),
# 当前 threshold=32(连续 32 次失败触发熔断),不在本脚本设置。 # 当前 threshold=32(连续 32 次失败触发熔断),不在本脚本设置。
@@ -58,4 +65,4 @@ CONCURRENCY="${CONCURRENCY:-12}"
# conda run --no-capture-output:不需 conda init(非交互脚本可跑),且实时直出不缓冲。 # conda run --no-capture-output:不需 conda init(非交互脚本可跑),且实时直出不缓冲。
conda run --no-capture-output -n Video-Tree-TRM python -u -m app.harness.video_split_cli \ conda run --no-capture-output -n Video-Tree-TRM python -u -m app.harness.video_split_cli \
--config config/video_split.yaml \ --config config/video_split.yaml \
--concurrency "${CONCURRENCY}" ${DRY_RUN_FLAG} --concurrency "${CONCURRENCY}" ${DRY_RUN_FLAG} ${FORCE_FLAG}