Upload folder using huggingface_hub

logs/finish_all/RESULTS.md

@@ -0,0 +1,48 @@
+# Corrected-cache retrain results -- 2026-09-03
+
+Models: run7' (base_c), arm B' (armB_c), arm C' (armC_c), arm A' (armA_c),
+all trained under YARN_ORIG_MAXPOS=4096 and evaluated under the same.
+Generative substring-match scoring, greedy, bf16, N=100 per cell.
+
+## PROMPT_LEN=1024
+```
+task,armA_c,armB_c,armC_c,base_c
+niah_single,89.0,87.0,92.0,87.0
+val_backward,2.0,2.0,1.0,3.0
+```
+
+## PROMPT_LEN=2048
+```
+task,armA_c,armB_c,armC_c,base_c
+niah_single,76.0,73.0,82.0,82.0
+val_backward,0.0,0.0,0.0,1.0
+```
+
+## PROMPT_LEN=4096
+```
+task,armA_c,armB_c,armC_c,base_c
+niah_single,62.0,60.0,68.0,71.0
+val_backward,2.0,1.0,3.0,4.0
+```
+
+## PROMPT_LEN=8192
+```
+task,armA_c,armB_c,armC_c,base_c
+niah_single,34.0,31.0,37.0,38.0
+val_backward,1.0,2.0,2.0,4.0
+```
+
+## PROMPT_LEN=16360
+```
+task,armA_c,armB_c,armC_c,base_c
+niah_single,25.0,23.0,26.0,27.0
+val_backward,4.0,1.0,0.0,0.0
+```
+
+## Training losses
+```
+run7'  (run 6, step 750): loss: 2.383854
+arm B' (run 7, step 400): loss: 2.775537
+arm c' : loss: 2.866560
+arm a' : loss: 2.550660
+```

logs/finish_all/arm_a.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2a0c6e92cb9aa17545267ed0e61f005bd18460ffc181c18b348bca6a6a3f1e33
+size 62

logs/finish_all/arm_c.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a299323ec48e000aaaf610226aa4fd25f5e618a3dd9705d72d83a019574852a5
+size 62

logs/finish_all/eval_1024.log

@@ -0,0 +1,101 @@
+[10:28:52] longeval: T=16384 prompt_len=1024 max_new=24 n=100 greedy=1 gpus=0,1,2,3
+[10:28:52] haystack: Data_Loader/Data/flux_val_000000.bin
+[10:28:52] fan-out: 4 GPU(s), NP=1 -> 8 concurrent job(s)
+[10:28:52]   base_c/niah_single already generated
+[10:28:52]   base_c/val_backward already generated
+[10:28:52]   armB_c/niah_single already generated
+[10:28:52]   armB_c/val_backward already generated
+[10:28:52] 4 job(s) to run
+[10:28:52]   [gpu 0] armC_c / niah_single
+[10:28:52]   [gpu 1] armC_c / val_backward
+[10:28:52]   [gpu 2] armA_c / niah_single
+[10:28:52]   [gpu 3] armA_c / val_backward
+[10:32:21]   scoring armC_c / niah_single
+
+=== niah_single  (prompt_len=1024, n=100) ===
+  accuracy: 92/100 = 92.0%
+    depth      acc   (needle position in context)
+     0.00  17/20   85.0%
+     0.25  17/20   85.0%
+     0.50  18/20   90.0%
+     0.75  20/20  100.0%
+     1.00  20/20  100.0%
+
+  --- 2 sample generations ---
+  [HIT ] expect=8056020
+         got= 8056020\n
+  [MISS] expect=5343902
+         got= a number that is not only unique to the film but also has a special meaning for the film.\n
+
+  wrote newck_len_1024/scores/armC_c_niah_single.csv
+[10:32:22]   scoring armC_c / val_backward
+
+=== val_backward  (prompt_len=1024, n=100) ===
+  accuracy: 1/100 = 1.0%
+    depth      acc   (needle position in context)
+     0.00   0/20    0.0%
+     0.25   0/20    0.0%
+     0.50   0/20    0.0%
+     0.75   0/20    0.0%
+     1.00   1/20    5.0%
+
+  --- 2 sample generations ---
+  [MISS] expect=yarrow
+         got= the\n3. I’m not sure if Moomin’s character is a dork, but
+  [MISS] expect=cactus
+         got= the\nscum (1979 - film version)\n
+
+  wrote newck_len_1024/scores/armC_c_val_backward.csv
+[10:32:22]   scoring armA_c / niah_single
+
+=== niah_single  (prompt_len=1024, n=100) ===
+  accuracy: 89/100 = 89.0%
+    depth      acc   (needle position in context)
+     0.00  17/20   85.0%
+     0.25  16/20   80.0%
+     0.50  17/20   85.0%
+     0.75  19/20   95.0%
+     1.00  20/20  100.0%
+
+  --- 2 sample generations ---
+  [HIT ] expect=8056020
+         got= 8056020\n
+  [MISS] expect=5343902
+         got= the number of times the cactus dies.\n
+
+  wrote newck_len_1024/scores/armA_c_niah_single.csv
+[10:32:22]   scoring armA_c / val_backward
+
+=== val_backward  (prompt_len=1024, n=100) ===
+  accuracy: 2/100 = 2.0%
+    depth      acc   (needle position in context)
+     0.00   0/20    0.0%
+     0.25   0/20    0.0%
+     0.50   0/20    0.0%
+     0.75   0/20    0.0%
+     1.00   2/20   10.0%
+
+  --- 2 sample generations ---
+  [MISS] expect=yarrow
+         got= a\n3. I’m not sure if Moomin is a fan of the game, but he�
+  [MISS] expect=cactus
+         got= the\nscum (1979 - film version)\n
+
+  wrote newck_len_1024/scores/armA_c_val_backward.csv
+[10:32:22] building summary
+
+=== per-task accuracy %  (secondary/descriptive: ~10 per depth cell) ===
+  task                 armA_c   armB_c   armC_c   base_c
+  niah_single            89.0     87.0     92.0     87.0
+  val_backward            2.0      2.0      1.0      3.0
+
+=== PRIMARY: accuracy % pooled by depth across tasks (+-95% Wilson) ===
+  depth             armA_c          armB_c          armC_c          base_c
+  0.00          42.5+-14.6      42.5+-14.6      42.5+-14.6      42.5+-14.6
+  0.25          40.0+-14.5      40.0+-14.5      42.5+-14.6      40.0+-14.5
+  0.50          42.5+-14.6      42.5+-14.6      45.0+-14.7      40.0+-14.5
+  0.75          47.5+-14.8      45.0+-14.7      50.0+-14.8      45.0+-14.7
+  1.00          55.0+-14.7      52.5+-14.8      52.5+-14.8      57.5+-14.6
+
+  wrote newck_len_1024/summary.csv and newck_len_1024/depth_summary.csv
+[10:32:22] DONE -> newck_len_1024/summary.csv

logs/finish_all/eval_16360.log

@@ -0,0 +1,101 @@
+[10:50:03] longeval: T=16384 prompt_len=16360 max_new=24 n=100 greedy=1 gpus=0,1,2,3
+[10:50:03] haystack: Data_Loader/Data/flux_val_000000.bin
+[10:50:03] fan-out: 4 GPU(s), NP=1 -> 8 concurrent job(s)
+[10:50:03]   base_c/niah_single already generated
+[10:50:03]   base_c/val_backward already generated
+[10:50:03]   armB_c/niah_single already generated
+[10:50:03]   armB_c/val_backward already generated
+[10:50:03] 4 job(s) to run
+[10:50:03]   [gpu 0] armC_c / niah_single
+[10:50:03]   [gpu 1] armC_c / val_backward
+[10:50:03]   [gpu 2] armA_c / niah_single
+[10:50:03]   [gpu 3] armA_c / val_backward
+[11:02:32]   scoring armC_c / niah_single
+
+=== niah_single  (prompt_len=16360, n=100) ===
+  accuracy: 26/100 = 26.0%
+    depth      acc   (needle position in context)
+     0.00   0/20    0.0%
+     0.25   0/20    0.0%
+     0.50   2/20   10.0%
+     0.75   7/20   35.0%
+     1.00  17/20   85.0%
+
+  --- 2 sample generations ---
+  [MISS] expect=8056020
+         got= the number of the moon.\n
+  [MISS] expect=5343902
+         got= 7.\nQuestion: What is the birth name of Anna Katarina?\n
+
+  wrote newck_len_16360/scores/armC_c_niah_single.csv
+[11:02:32]   scoring armC_c / val_backward
+
+=== val_backward  (prompt_len=16360, n=100) ===
+  accuracy: 0/100 = 0.0%
+    depth      acc   (needle position in context)
+     0.00   0/20    0.0%
+     0.25   0/20    0.0%
+     0.50   0/20    0.0%
+     0.75   0/20    0.0%
+     1.00   0/20    0.0%
+
+  --- 2 sample generations ---
+  [MISS] expect=yarrow
+         got= the one that has the special magic number 8056020.\n
+  [MISS] expect=cactus
+         got= called "The Magic Number".\n
+
+  wrote newck_len_16360/scores/armC_c_val_backward.csv
+[11:02:32]   scoring armA_c / niah_single
+
+=== niah_single  (prompt_len=16360, n=100) ===
+  accuracy: 25/100 = 25.0%
+    depth      acc   (needle position in context)
+     0.00   0/20    0.0%
+     0.25   0/20    0.0%
+     0.50   2/20   10.0%
+     0.75   6/20   30.0%
+     1.00  17/20   85.0%
+
+  --- 2 sample generations ---
+  [MISS] expect=8056020
+         got= the number of the moon.\n
+  [MISS] expect=5343902
+         got= 7.\nWhat is Anna Katarina's birthday?\n
+
+  wrote newck_len_16360/scores/armA_c_niah_single.csv
+[11:02:33]   scoring armA_c / val_backward
+
+=== val_backward  (prompt_len=16360, n=100) ===
+  accuracy: 4/100 = 4.0%
+    depth      acc   (needle position in context)
+     0.00   0/20    0.0%
+     0.25   0/20    0.0%
+     0.50   0/20    0.0%
+     0.75   0/20    0.0%
+     1.00   4/20   20.0%
+
+  --- 2 sample generations ---
+  [MISS] expect=yarrow
+         got= the one that has the special magic number 8056020.\n
+  [MISS] expect=cactus
+         got=: "The Magician".\n
+
+  wrote newck_len_16360/scores/armA_c_val_backward.csv
+[11:02:33] building summary
+
+=== per-task accuracy %  (secondary/descriptive: ~10 per depth cell) ===
+  task                 armA_c   armB_c   armC_c   base_c
+  niah_single            25.0     23.0     26.0     27.0
+  val_backward            4.0      1.0      0.0      0.0
+
+=== PRIMARY: accuracy % pooled by depth across tasks (+-95% Wilson) ===
+  depth             armA_c          armB_c          armC_c          base_c
+  0.00           0.0+- 4.4       0.0+- 4.4       0.0+- 4.4       0.0+- 4.4
+  0.25           0.0+- 4.4       0.0+- 4.4       0.0+- 4.4       0.0+- 4.4
+  0.50           5.0+- 7.6       5.0+- 7.6       5.0+- 7.6       5.0+- 7.6
+  0.75          15.0+-11.0      10.0+- 9.5      17.5+-11.6      17.5+-11.6
+  1.00          52.5+-14.8      45.0+-14.7      42.5+-14.6      45.0+-14.7
+
+  wrote newck_len_16360/summary.csv and newck_len_16360/depth_summary.csv
+[11:02:33] DONE -> newck_len_16360/summary.csv

logs/finish_all/eval_2048.log

@@ -0,0 +1,103 @@
+[10:32:22] longeval: T=16384 prompt_len=2048 max_new=24 n=100 greedy=1 gpus=0,1,2,3
+[10:32:22] haystack: Data_Loader/Data/flux_val_000000.bin
+[10:32:22] fan-out: 4 GPU(s), NP=1 -> 8 concurrent job(s)
+[10:32:23]   base_c/niah_single already generated
+[10:32:23]   base_c/val_backward already generated
+[10:32:23]   armB_c/niah_single already generated
+[10:32:23]   armB_c/val_backward already generated
+[10:32:23] 4 job(s) to run
+[10:32:23]   [gpu 0] armC_c / niah_single
+[10:32:23]   [gpu 1] armC_c / val_backward
+[10:32:23]   [gpu 2] armA_c / niah_single
+[10:32:23]   [gpu 3] armA_c / val_backward
+[10:36:19]   scoring armC_c / niah_single
+
+=== niah_single  (prompt_len=2048, n=100) ===
+  accuracy: 82/100 = 82.0%
+    depth      acc   (needle position in context)
+     0.00  15/20   75.0%
+     0.25  16/20   80.0%
+     0.50  17/20   85.0%
+     0.75  17/20   85.0%
+     1.00  17/20   85.0%
+
+  --- 2 sample generations ---
+  [MISS] expect=8056020
+         got= 7.\nThe special magic number for yarrow is 7.\n
+  [MISS] expect=5343902
+         got= 2.\nQuestion: How does the magic number for cactus work?\n
+
+  wrote newck_len_2048/scores/armC_c_niah_single.csv
+[10:36:20]   scoring armC_c / val_backward
+
+=== val_backward  (prompt_len=2048, n=100) ===
+  accuracy: 0/100 = 0.0%
+    depth      acc   (needle position in context)
+     0.00   0/20    0.0%
+     0.25   0/20    0.0%
+     0.50   0/20    0.0%
+     0.75   0/20    0.0%
+     1.00   0/20    0.0%
+
+  --- 2 sample generations ---
+  [MISS] expect=yarrow
+         got= the\nscum (1979 - film version)\n
+  [MISS] expect=cactus
+         got= the Magic Number 5343902. It is a unique number that is used to unlock special abilities in various magic-
+
+  wrote newck_len_2048/scores/armC_c_val_backward.csv
+[10:36:20]   scoring armA_c / niah_single
+
+=== niah_single  (prompt_len=2048, n=100) ===
+  accuracy: 76/100 = 76.0%
+    depth      acc   (needle position in context)
+     0.00  11/20   55.0%
+     0.25  15/20   75.0%
+     0.50  18/20   90.0%
+     0.75  17/20   85.0%
+     1.00  15/20   75.0%
+
+  --- 2 sample generations ---
+  [MISS] expect=8056020
+         got= the number of the sun.\n
+  [MISS] expect=5343902
+         got= 2.\nQuestion: How do I find a reliable cactus rental company?\n
+
+  wrote newck_len_2048/scores/armA_c_niah_single.csv
+[10:36:20]   scoring armA_c / val_backward
+
+=== val_backward  (prompt_len=2048, n=100) ===
+  accuracy: 0/100 = 0.0%
+    depth      acc   (needle position in context)
+     0.00   0/20    0.0%
+     0.25   0/20    0.0%
+     0.50   0/20    0.0%
+     0.75   0/20    0.0%
+     1.00   0/20    0.0%
+
+  --- 2 sample generations ---
+  [MISS] expect=yarrow
+         got= the\nscum (1979 - film version)\n
+  [MISS] expect=cactus
+         got= the Magic Wheel.\n
+
+  wrote newck_len_2048/scores/armA_c_val_backward.csv
+[10:36:20] building summary
+
+=== per-task accuracy %  (secondary/descriptive: ~10 per depth cell) ===
+  task                 armA_c   armB_c   armC_c   base_c
+  niah_single            76.0     73.0     82.0     82.0
+  val_backward            0.0      0.0      0.0      1.0
+
+  degenerate (<=1% for every model): ['val_backward']
+
+=== PRIMARY: accuracy % pooled by depth across tasks (+-95% Wilson) ===
+  depth             armA_c          armB_c          armC_c          base_c
+  0.00          55.0+-20.0      40.0+-19.7      75.0+-17.8      55.0+-20.0
+  0.25          75.0+-17.8      75.0+-17.8      80.0+-16.8      70.0+-18.7
+  0.50          90.0+-13.7      90.0+-13.7      85.0+-15.4      95.0+-11.4
+  0.75          85.0+-15.4      80.0+-16.8      85.0+-15.4      95.0+-11.4
+  1.00          75.0+-17.8      80.0+-16.8      85.0+-15.4      95.0+-11.4
+
+  wrote newck_len_2048/summary.csv and newck_len_2048/depth_summary.csv
+[10:36:20] DONE -> newck_len_2048/summary.csv

logs/finish_all/eval_4096.log

@@ -0,0 +1,101 @@
+[10:36:20] longeval: T=16384 prompt_len=4096 max_new=24 n=100 greedy=1 gpus=0,1,2,3
+[10:36:20] haystack: Data_Loader/Data/flux_val_000000.bin
+[10:36:20] fan-out: 4 GPU(s), NP=1 -> 8 concurrent job(s)
+[10:36:20]   base_c/niah_single already generated
+[10:36:20]   base_c/val_backward already generated
+[10:36:20]   armB_c/niah_single already generated
+[10:36:20]   armB_c/val_backward already generated
+[10:36:20] 4 job(s) to run
+[10:36:20]   [gpu 0] armC_c / niah_single
+[10:36:20]   [gpu 1] armC_c / val_backward
+[10:36:20]   [gpu 2] armA_c / niah_single
+[10:36:20]   [gpu 3] armA_c / val_backward
+[10:41:34]   scoring armC_c / niah_single
+
+=== niah_single  (prompt_len=4096, n=100) ===
+  accuracy: 68/100 = 68.0%
+    depth      acc   (needle position in context)
+     0.00  10/20   50.0%
+     0.25  10/20   50.0%
+     0.50  13/20   65.0%
+     0.75  17/20   85.0%
+     1.00  18/20   90.0%
+
+  --- 2 sample generations ---
+  [HIT ] expect=8056020
+         got= 8056020.\n
+  [MISS] expect=5343902
+         got= cactus.\n
+
+  wrote newck_len_4096/scores/armC_c_niah_single.csv
+[10:41:35]   scoring armC_c / val_backward
+
+=== val_backward  (prompt_len=4096, n=100) ===
+  accuracy: 3/100 = 3.0%
+    depth      acc   (needle position in context)
+     0.00   0/20    0.0%
+     0.25   0/20    0.0%
+     0.50   0/20    0.0%
+     0.75   0/20    0.0%
+     1.00   3/20   15.0%
+
+  --- 2 sample generations ---
+  [MISS] expect=yarrow
+         got= the number 8056020. It is a special number that is used to identify the bus.\n
+  [MISS] expect=cactus
+         got= the one that has the special magic number 5343902.\n
+
+  wrote newck_len_4096/scores/armC_c_val_backward.csv
+[10:41:35]   scoring armA_c / niah_single
+
+=== niah_single  (prompt_len=4096, n=100) ===
+  accuracy: 62/100 = 62.0%
+    depth      acc   (needle position in context)
+     0.00   7/20   35.0%
+     0.25   9/20   45.0%
+     0.50  10/20   50.0%
+     0.75  17/20   85.0%
+     1.00  19/20   95.0%
+
+  --- 2 sample generations ---
+  [MISS] expect=8056020
+         got= 8.\nQuestion: How do I find a reliable bus rental company?\n
+  [MISS] expect=5343902
+         got= cactus.\n
+
+  wrote newck_len_4096/scores/armA_c_niah_single.csv
+[10:41:35]   scoring armA_c / val_backward
+
+=== val_backward  (prompt_len=4096, n=100) ===
+  accuracy: 2/100 = 2.0%
+    depth      acc   (needle position in context)
+     0.00   0/20    0.0%
+     0.25   0/20    0.0%
+     0.50   0/20    0.0%
+     0.75   0/20    0.0%
+     1.00   2/20   10.0%
+
+  --- 2 sample generations ---
+  [MISS] expect=yarrow
+         got= the “8” symbol.\n
+  [MISS] expect=cactus
+         got= the one that has the special magic number 5343902.\n
+
+  wrote newck_len_4096/scores/armA_c_val_backward.csv
+[10:41:35] building summary
+
+=== per-task accuracy %  (secondary/descriptive: ~10 per depth cell) ===
+  task                 armA_c   armB_c   armC_c   base_c
+  niah_single            62.0     60.0     68.0     71.0
+  val_backward            2.0      1.0      3.0      4.0
+
+=== PRIMARY: accuracy % pooled by depth across tasks (+-95% Wilson) ===
+  depth             armA_c          armB_c          armC_c          base_c
+  0.00          17.5+-11.6      17.5+-11.6      25.0+-13.0      22.5+-12.6
+  0.25          22.5+-12.6      20.0+-12.1      25.0+-13.0      27.5+-13.4
+  0.50          25.0+-13.0      25.0+-13.0      32.5+-13.9      35.0+-14.2
+  0.75          42.5+-14.6      40.0+-14.5      42.5+-14.6      42.5+-14.6
+  1.00          52.5+-14.8      50.0+-14.8      52.5+-14.8      60.0+-14.5
+
+  wrote newck_len_4096/summary.csv and newck_len_4096/depth_summary.csv
+[10:41:35] DONE -> newck_len_4096/summary.csv

logs/finish_all/eval_8192.log

@@ -0,0 +1,101 @@
+[10:41:35] longeval: T=16384 prompt_len=8192 max_new=24 n=100 greedy=1 gpus=0,1,2,3
+[10:41:35] haystack: Data_Loader/Data/flux_val_000000.bin
+[10:41:35] fan-out: 4 GPU(s), NP=1 -> 8 concurrent job(s)
+[10:41:35]   base_c/niah_single already generated
+[10:41:35]   base_c/val_backward already generated
+[10:41:35]   armB_c/niah_single already generated
+[10:41:35]   armB_c/val_backward already generated
+[10:41:35] 4 job(s) to run
+[10:41:35]   [gpu 0] armC_c / niah_single
+[10:41:35]   [gpu 1] armC_c / val_backward
+[10:41:35]   [gpu 2] armA_c / niah_single
+[10:41:35]   [gpu 3] armA_c / val_backward
+[10:50:02]   scoring armC_c / niah_single
+
+=== niah_single  (prompt_len=8192, n=100) ===
+  accuracy: 37/100 = 37.0%
+    depth      acc   (needle position in context)
+     0.00   2/20   10.0%
+     0.25   4/20   20.0%
+     0.50   4/20   20.0%
+     0.75  10/20   50.0%
+     1.00  17/20   85.0%
+
+  --- 2 sample generations ---
+  [MISS] expect=8056020
+         got= 0.\nunpx[asciilifeform]: I'm not sure what it is, but I'm sure
+  [MISS] expect=5343902
+         got= the number of the plant’s leaves.\n
+
+  wrote newck_len_8192/scores/armC_c_niah_single.csv
+[10:50:03]   scoring armC_c / val_backward
+
+=== val_backward  (prompt_len=8192, n=100) ===
+  accuracy: 2/100 = 2.0%
+    depth      acc   (needle position in context)
+     0.00   0/20    0.0%
+     0.25   0/20    0.0%
+     0.50   0/20    0.0%
+     0.75   0/20    0.0%
+     1.00   2/20   10.0%
+
+  --- 2 sample generations ---
+  [MISS] expect=yarrow
+         got= the one that has the special magic number 8056020.\n
+  [MISS] expect=cactus
+         got= the number 5343902.\n
+
+  wrote newck_len_8192/scores/armC_c_val_backward.csv
+[10:50:03]   scoring armA_c / niah_single
+
+=== niah_single  (prompt_len=8192, n=100) ===
+  accuracy: 34/100 = 34.0%
+    depth      acc   (needle position in context)
+     0.00   2/20   10.0%
+     0.25   2/20   10.0%
+     0.50   4/20   20.0%
+     0.75  10/20   50.0%
+     1.00  16/20   80.0%
+
+  --- 2 sample generations ---
+  [MISS] expect=8056020
+         got= 1.\nuncolorful: I don't know what it is, but I'm sure it's a random number
+  [MISS] expect=5343902
+         got= the number of the plant’s leaves.\n
+
+  wrote newck_len_8192/scores/armA_c_niah_single.csv
+[10:50:03]   scoring armA_c / val_backward
+
+=== val_backward  (prompt_len=8192, n=100) ===
+  accuracy: 1/100 = 1.0%
+    depth      acc   (needle position in context)
+     0.00   0/20    0.0%
+     0.25   0/20    0.0%
+     0.50   0/20    0.0%
+     0.75   0/20    0.0%
+     1.00   1/20    5.0%
+
+  --- 2 sample generations ---
+  [MISS] expect=yarrow
+         got= the one that has the special magic number 8056020.\n
+  [MISS] expect=cactus
+         got= the one that has the special magic number 5343902.\n
+
+  wrote newck_len_8192/scores/armA_c_val_backward.csv
+[10:50:03] building summary
+
+=== per-task accuracy %  (secondary/descriptive: ~10 per depth cell) ===
+  task                 armA_c   armB_c   armC_c   base_c
+  niah_single            34.0     31.0     37.0     38.0
+  val_backward            1.0      2.0      2.0      4.0
+
+=== PRIMARY: accuracy % pooled by depth across tasks (+-95% Wilson) ===
+  depth             armA_c          armB_c          armC_c          base_c
+  0.00           5.0+- 7.6       7.5+- 8.6       5.0+- 7.6       7.5+- 8.6
+  0.25           5.0+- 7.6       5.0+- 7.6      10.0+- 9.5       5.0+- 7.6
+  0.50          10.0+- 9.5      10.0+- 9.5      10.0+- 9.5      12.5+-10.3
+  0.75          25.0+-13.0      17.5+-11.6      25.0+-13.0      25.0+-13.0
+  1.00          42.5+-14.6      42.5+-14.6      47.5+-14.8      55.0+-14.7
+
+  wrote newck_len_8192/summary.csv and newck_len_8192/depth_summary.csv
+[10:50:03] DONE -> newck_len_8192/summary.csv

logs/finish_all/finish.log

@@ -0,0 +1,40 @@
+[09:02:29] === finish_all.sh start ===
+[09:02:29] free disk 71G
+[09:02:29] WARNING: no HF token found. Training and eval will proceed; the UPLOAD
+[09:02:29]   step will wait for one. Provide it either way:
+[09:02:29]     echo '<token>' > /root/Adhi/BluTrain/dist/Context_Parallelism/hf_token.txt
+[09:02:29]     or: export HF_TOKEN=<token>  (only helps if set before this script)
+[09:04:06] === finish_all.sh start ===
+[09:04:06] free disk 71G
+[09:04:06] WARNING: no HF token found. Training and eval will proceed; the UPLOAD
+[09:04:06]   step will wait for one. Provide it either way:
+[09:04:06]     echo '<token>' > /root/Adhi/BluTrain/dist/Context_Parallelism/hf_token.txt
+[09:04:06]     or: export HF_TOKEN=<token>  (only helps if set before this script)
+[09:04:06] training arm c' (400 steps from run7')
+[09:46:27] arm c' -> run 27, blumodelcp_run27_step_400.ckpt, final loss loss: 2.866560
+[09:46:27] training arm a' (400 steps from run7')
+[10:28:52] arm a' -> run 28, blumodelcp_run28_step_400.ckpt, final loss loss: 2.550660
+[10:28:52] eval models: base_c armB_c armC_c armA_c 
+[10:28:52] eval PROMPT_LEN=1024 (finished cells are skipped)
+[10:32:22]   -> task,armA_c,armB_c,armC_c,base_c
+[10:32:22] eval PROMPT_LEN=2048 (finished cells are skipped)
+[10:36:20]   -> task,armA_c,armB_c,armC_c,base_c
+[10:36:20] eval PROMPT_LEN=4096 (finished cells are skipped)
+[10:41:35]   -> task,armA_c,armB_c,armC_c,base_c
+[10:41:35] eval PROMPT_LEN=8192 (finished cells are skipped)
+[10:50:03]   -> task,armA_c,armB_c,armC_c,base_c
+[10:50:03] eval PROMPT_LEN=16360 (finished cells are skipped)
+[11:02:33]   -> task,armA_c,armB_c,armC_c,base_c
+[11:02:33] eval done
+[11:02:35] notebook rebuilt: cache_fix_results.ipynb
+[11:02:35] wrote finish_logs/RESULTS.md
+[11:02:35] uploading to unparallelled/BluScriptCP
+[11:02:36]   ok   cache_fix_results.ipynb -> docs/cache_fix_results_2026-09-03.ipynb
+[11:02:38]   ok   finish_logs/RESULTS.md -> docs/RETRAIN_RESULTS_2026-09-03.md
+[11:02:40]   ok   EVAL_COMPARISON.md -> docs/EVAL_COMPARISON_2026-09-02.md
+[11:02:44]   ok   newck_len_1024 -> logs/newck_len_1024/
+[11:02:48]   ok   newck_len_2048 -> logs/newck_len_2048/
+[11:02:52]   ok   newck_len_4096 -> logs/newck_len_4096/
+[11:02:56]   ok   newck_len_8192 -> logs/newck_len_8192/
+[11:03:00]   ok   newck_len_16360 -> logs/newck_len_16360/
+[11:03:02]   ok   retrain_logs -> logs/retrain_corrected_cache/

logs/finish_all/notebook.log

@@ -0,0 +1,6 @@
+[IPKernelApp] WARNING | Kernel is running over TCP without encryption. All communication (including code and outputs) is sent in plain text and is susceptible to eavesdropping. Use IPC transport or launch with kernel manager-provisioned CurveZMQ keys to enable transport encryption.
+  old matched  : [1024, 2048, 4096, 8192, 16360]
+  old mismatch : [1024, 2048, 4096, 8192, 16360]
+  new matched  : [1024, 2048, 4096, 8192, 16360]
+  executing...
+  wrote cache_fix_results.ipynb (326716 B, 15 cells)

logs/finish_all/upload.log

@@ -0,0 +1,23 @@
+url=https://huggingface.co/unparallelled/BluScriptCP/commit/8979a4757c9ab85ff95880598c4671b3fcce2c07
+url=https://huggingface.co/unparallelled/BluScriptCP/commit/2fb5680048cb7220d12bbdf893d56259cc7829cc
+url=https://huggingface.co/unparallelled/BluScriptCP/commit/31557bff0e606ac6678a9b95c8dce649b2e5b895
+Found 48 files to upload
+Uploading... 48/48 files checked, 0/2 uploaded (0.00B transferred), 0 committed in 0 commit(s)
+url=https://huggingface.co/unparallelled/BluScriptCP/commit/db3b3c178868cb395bfea986ae10faedb57865e9
+Found 48 files to upload
+Uploading... 48/48 files checked, 0/2 uploaded (0.00B transferred), 0 committed in 0 commit(s)
+url=https://huggingface.co/unparallelled/BluScriptCP/commit/74aecb7af91c76ad5683dd4b82dc79c34b4be63a
+Found 48 files to upload
+Uploading... 48/48 files checked, 0/0 uploaded (0.00B transferred), 0 committed in 0 commit(s)
+url=https://huggingface.co/unparallelled/BluScriptCP/commit/40f29a5268dee593898f217ccb789e5261e2fef6
+Found 48 files to upload
+Uploading... 48/48 files checked, 0/2 uploaded (0.00B transferred), 0 committed in 0 commit(s)
+url=https://huggingface.co/unparallelled/BluScriptCP/commit/aa57c4c988641b965ed7bc2297b9614749f8dcae
+Found 48 files to upload
+Uploading... 48/48 files checked, 0/2 uploaded (0.00B transferred), 0 committed in 0 commit(s)
+url=https://huggingface.co/unparallelled/BluScriptCP/commit/4a9a7da39046b2452ce5ec674d63e5b3b6568160
+Found 7 files to upload
+Uploading... 7/7 files checked, 0/0 uploaded (0.00B transferred), 0 committed in 0 commit(s)
+url=https://huggingface.co/unparallelled/BluScriptCP/commit/c518028f1442415c8e67ab1c52a5400cca541472
+Found 13 files to upload
+Uploading... 13/13 files checked, 0/0 uploaded (0.00B transferred), 0 committed in 0 commit(s)