exp(dagger-deploy): 轮次21——DAgger-3(部署配置采集500题+15000步微调)负结论:部署口径 39.5% vs dag2 43.8%,开环退化到19.5mm;下一步试冻结编码器只调解码器

This commit is contained in:
cbb
2026-10-20 09:41:44 +08:00
parent bb919d33c9
commit 7a8e5aab63
4 changed files with 33 additions and 2 deletions
+1 -1
View File
@@ -24,7 +24,7 @@ for ((r=1; r<=R; r++)); do
SAMPLES="$SAMPLES $SD"
if [ -f $OUT/done ]; then echo "[dagger] 第 $r 轮已完成,跳过"; CK=$($PY -m ml.common.ckpt pick $OUT); continue; fi
if [ -f $D/collect.json ]; then echo "[dagger] 第 $r 轮起点已采集,跳过"
else echo "[dagger] 第 $r 轮:用 $CK 闭环采起点($N 题)"; $PY -m ml.dagger.collect --ckpt $CK --samples ml/data/samples --n $N --seed $r --parts $P --out $D --fb_gain ${FB_GAIN:-0} --safety ${DAG_SAFETY:-0}; fi
else echo "[dagger] 第 $r 轮:用 $CK 闭环采起点($N 题)"; $PY -m ml.dagger.collect --ckpt $CK --samples ml/data/samples --n $N --seed $r --parts $P --out $D --fb_gain ${FB_GAIN:-0} --safety ${DAG_SAFETY:-0} --replan ${DAG_REPLAN:-r3} --commit_n ${DAG_COMMIT_N:-3} --safety_partial ${DAG_PARTIAL:-0}; fi
cat $D/collect.json
pids=()
for f in $D/tasks_*.jsonl; do