[18:22:46] longeval: T=65536 prompt_len=65512 max_new=24 n=100 greedy=1 gpus=3
[18:22:46] haystack: /root/Adhi/BluTrain/dist/Context_Parallelism/Data_Loader/Data/b2_8192_16384/val_shard_00000.bin
[18:22:46] prompt signature changed -- regenerating prompts and clearing state
[18:22:46] generating prompts: niah_single
[18:22:51] fan-out: 1 GPU(s), NP=1 -> 1 concurrent job(s)
[18:22:51] 1 job(s) to run
[18:22:51]   [gpu 3] cont64k / niah_single
[19:07:25]   scoring cont64k / niah_single

=== niah_single  (prompt_len=65512, 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=8056020
         got= the number yof. Let’ credentials jacking wheel. IB Text Is to answer, as does
  [MISS] expect=5343902
         got= not just what the output. Why do it better Insaid had two options and just the questions of what happens, when

  wrote cont64k_len_65512/scores/cont64k_niah_single.csv
[19:07:25] building summary

=== per-task accuracy %  (secondary/descriptive: ~10 per depth cell) ===
  task                cont64k
  niah_single             1.0

  degenerate (<=1% for every model): ['niah_single']
  ALL tasks degenerate -> FLOOR EFFECT. Pooling everything anyway;
  differences between models are untestable at this setting (no power).

=== PRIMARY: accuracy % pooled by depth across tasks (+-95% Wilson) ===
  depth            cont64k
  0.00           0.0+- 8.1
  0.25           0.0+- 8.1
  0.50           0.0+- 8.1
  0.75           0.0+- 8.1
  1.00           5.0+-11.4

  wrote cont64k_len_65512/summary.csv and cont64k_len_65512/depth_summary.csv
[19:07:25] DONE -> cont64k_len_65512/summary.csv