Add files using upload-large-folder tool

runs_laq_lt/eval/cocolt_2ep_hires/summary.txt

@@ -0,0 +1,112 @@
+==============================================================================
+EVAL SUMMARY  |  cocolt_2ep_hires
+==============================================================================
+
+Metric: mean-over-IoU (fastevaluate). BPS = boxes/second, batch 1, 1 GPU.
+fallback% = switch_to_ar / forward_steps  (fraction of block-decisions; paper implies ~7-8%)
+
+      profile    mode |     F1      P      R  F1@.5 |    BPS  steps  boxes fallb%
+---------------------------------------------------------------------------------
+  paper_seed0    slow |   54.5   57.1   52.2   73.4 |   2.70   67.3   7.05    0.0
+  paper_seed0    fast |   53.1   63.0   45.8   68.2 |   8.91   15.2   6.95    0.0
+  paper_seed0  hybrid |   54.2   63.6   47.1   69.6 |   8.05   18.3   7.06    4.6
+
+  paper_seed1    slow |   54.5   57.0   52.3   73.0 |   2.71   67.3   7.04    0.0
+  paper_seed1    fast |   53.0   63.0   45.8   68.2 |   8.90   15.3   6.96    0.0
+  paper_seed1  hybrid |   54.0   63.5   46.9   69.3 |   8.05   18.2   7.05    4.5
+
+ paper_greedy    slow |   56.8   58.9   54.9   74.0 |   2.78   70.4   7.38    0.0
+ paper_greedy    fast |   55.2   59.8   51.2   72.0 |   9.17   15.2   7.07    0.0
+ paper_greedy  hybrid |   56.2   60.2   52.6   73.1 |   8.34   18.7   7.21    5.1
+
+       legacy    slow |   56.8   58.9   54.9   74.0 |   2.76   70.4   7.38    0.0
+       legacy    fast |   55.2   59.8   51.2   71.9 |   9.23   15.2   7.07    0.0
+       legacy  hybrid |   56.1   60.2   52.6   73.1 |   8.32   18.7   7.21    5.1
+
+paper is STOCHASTIC (nucleus sampling). Mean +/- half-spread over 2 seeds:
+---------------------------------------------------------------------------------
+ paper (mean)    slow | F1   54.5+/-0.0   BPS    2.7+/-0.0   fallback%    0.0+/-0.0
+ paper (mean)    fast | F1   53.0+/-0.0   BPS    8.9+/-0.0   fallback%    0.0+/-0.0
+ paper (mean)  hybrid | F1   54.1+/-0.1   BPS    8.1+/-0.0   fallback%    4.6+/-0.0
+
+  -> a gap smaller than the spread is NOT evidence. Use paper_greedy
+     (deterministic) for tight comparisons.
+
+INTEGRITY  legacy/slow (56.83) vs paper_greedy/slow (56.83) -> OK (identical, as required)
+
+B1 READ-OUT  hybrid fallback rate (switch_to_ar/forward_steps): legacy 5.1%  vs  paper 4.6%   (paper ~7-8%; the RELATIVE legacy>paper gap is the primary signal)
+  The 0.9 abnormality threshold is calibrated on the SHARPENED (T=0.7 +
+  top-p) distribution. Greedy's flatter softmax trips it far more often, so
+  legacy should be markedly higher. If they match, B1 has not taken effect.
+
+==============================================================================
+PROMPT-SHAPE (n_cats) OOD  --  profile: paper_greedy
+==============================================================================
+Training chunks each image's categories into queries of 1..3. The eval set
+asks for the image's FULL list, and 28.8% of eval queries have >3 categories
+-- a prompt shape the model never saw. Caveat: images with more categories
+also have more objects, so a drop may be difficulty, not novelty.
+
+--- paper_greedy/slow ---
+predictions: 4952  |  eval queries: 4952  |  unbucketed: 0
+
+    n_cats |     N |    F1 |     P |     R | F1@.5
+--------------------------------------------------
+         1 |  1025 |  74.1 |  78.2 |  70.3 |  87.2
+         2 |  1522 |  67.8 |  69.2 |  66.5 |  82.9
+         3 |   980 |  59.8 |  63.4 |  56.5 |  76.5
+       4-5 |   935 |  55.5 |  58.8 |  52.5 |  73.0
+        6+ |   490 |  51.7 |  56.0 |  48.0 |  70.5
+
+    n_cats |     N |    F1 |     P |     R | F1@.5
+--------------------------------------------------
+  <=3 (IN) |  3527 |  62.9 |  64.5 |  61.3 |  78.4
+  >3 (OOD) |  1425 |  52.7 |  56.1 |  49.7 |  70.7
+
+If '>3 (OOD)' is materially below '<=3 (IN)', the prompt-shape gap
+is real: training never showed the model a 4+-category query.
+Careful: bigger images have more categories AND more objects, so a
+small drop may just be task difficulty, not prompt-shape OOD.
+
+--- paper_greedy/fast ---
+predictions: 4952  |  eval queries: 4952  |  unbucketed: 0
+
+    n_cats |     N |    F1 |     P |     R | F1@.5
+--------------------------------------------------
+         1 |  1025 |  72.1 |  76.8 |  68.0 |  85.1
+         2 |  1522 |  65.6 |  69.2 |  62.3 |  80.8
+         3 |   980 |  57.6 |  63.0 |  53.0 |  73.8
+       4-5 |   935 |  54.5 |  61.9 |  48.7 |  71.4
+        6+ |   490 |  49.1 |  57.4 |  42.9 |  66.0
+
+    n_cats |     N |    F1 |     P |     R | F1@.5
+--------------------------------------------------
+  <=3 (IN) |  3527 |  61.1 |  65.2 |  57.4 |  76.5
+  >3 (OOD) |  1425 |  51.0 |  58.2 |  45.4 |  67.8
+
+If '>3 (OOD)' is materially below '<=3 (IN)', the prompt-shape gap
+is real: training never showed the model a 4+-category query.
+Careful: bigger images have more categories AND more objects, so a
+small drop may just be task difficulty, not prompt-shape OOD.
+
+--- paper_greedy/hybrid ---
+predictions: 4952  |  eval queries: 4952  |  unbucketed: 0
+
+    n_cats |     N |    F1 |     P |     R | F1@.5
+--------------------------------------------------
+         1 |  1025 |  74.1 |  78.4 |  70.2 |  87.0
+         2 |  1522 |  66.8 |  69.2 |  64.6 |  82.0
+         3 |   980 |  58.6 |  64.0 |  54.0 |  75.0
+       4-5 |   935 |  55.1 |  62.4 |  49.3 |  72.4
+        6+ |   490 |  50.0 |  57.4 |  44.3 |  67.4
+
+    n_cats |     N |    F1 |     P |     R | F1@.5
+--------------------------------------------------
+  <=3 (IN) |  3527 |  62.5 |  65.8 |  59.4 |  78.2
+  >3 (OOD) |  1425 |  51.8 |  58.7 |  46.3 |  68.9
+
+If '>3 (OOD)' is materially below '<=3 (IN)', the prompt-shape gap
+is real: training never showed the model a 4+-category query.
+Careful: bigger images have more categories AND more objects, so a
+small drop may just be task difficulty, not prompt-shape OOD.
+