Upload folder using huggingface_hub

logs/retrain_corrected_cache/arm_a.ckptpath

@@ -0,0 +1 @@
+DEFERRED-not-trained (arm A intentionally skipped 2026-09-03; run it later with: rm retrain_logs/arm_a.ckptpath && ./retrain_corrected_cache.sh)

logs/retrain_corrected_cache/arm_b.runnum

@@ -0,0 +1 @@
+7

logs/retrain_corrected_cache/driver.out

@@ -0,0 +1,5 @@
+[07:19:19] free disk 92G (need ~70G: 4 checkpoint sets at ~17G)
+[07:19:19] arm data ready
+[07:19:19] run7' already done (checkpoints_bluscriptcp/rank_0/blumodelcp_run6_step_750.ckpt)
+[07:19:19] arm a already done
+[07:19:19] training arm b (corrected cache)

logs/retrain_corrected_cache/master.log

@@ -0,0 +1,31 @@
+[04:40:01] free disk 116G (need ~70G: 4 checkpoint sets at ~17G)
+[04:40:01] generating arm a data
+wrote 3 shard(s) in in2_data/arm_a  209977344 tokens total
+  answer tokens 0 (0.00%)  synthetic 0 (0.00%)
+wrote 1 shard(s) in in2_data/arm_a  1048576 tokens total
+  answer tokens 0 (0.00%)  synthetic 0 (0.00%)
+[04:40:09] generating arm b data
+wrote 3 shard(s) in in2_data/arm_b  209977344 tokens total
+  answer tokens 6561792 (3.12%)  synthetic 22146048 (10.55%)
+  distance: min 12  median 12  max 12  n=820224
+wrote 1 shard(s) in in2_data/arm_b  1048576 tokens total
+  answer tokens 32768 (3.12%)  synthetic 110592 (10.55%)
+  distance: min 12  median 12  max 12  n=4096
+[04:40:37] generating arm c data
+wrote 3 shard(s) in in2_data/arm_c  209977344 tokens total
+  answer tokens 6561792 (3.12%)  synthetic 22146048 (10.55%)
+  distance: min 76  median 2923  max 16350  n=820224
+wrote 1 shard(s) in in2_data/arm_c  1048576 tokens total
+  answer tokens 32768 (3.12%)  synthetic 110592 (10.55%)
+  distance: min 76  median 2921  max 16194  n=4096
+[04:41:05] arm data ready
+[04:41:05] training run7' (YaRN 16k extension, corrected cache)
+[07:14:12]   run7' (YaRN 16k extension, corrected cache) -> run 6, blumodelcp_run6_step_750.ckpt, final loss loss: 2.383854
+[07:14:12] arm a already done
+[07:14:12] training arm b (corrected cache)
+[07:14:59] FATAL: arm b (corrected cache) training failed -- see retrain_logs/arm_b.log
+[07:19:19] free disk 92G (need ~70G: 4 checkpoint sets at ~17G)
+[07:19:19] arm data ready
+[07:19:19] run7' already done (checkpoints_bluscriptcp/rank_0/blumodelcp_run6_step_750.ckpt)
+[07:19:19] arm a already done
+[07:19:19] training arm b (corrected cache)

logs/retrain_corrected_cache/run7c.ckptpath

@@ -0,0 +1 @@
+checkpoints_bluscriptcp/rank_0/blumodelcp_run6_step_750.ckpt