colab2 final (c2s2200): CE curve + teacher parity logs

colab2/eval/colab2_ce_curve.txt

@@ -0,0 +1,12 @@
+[10:39:32 + 15.5m]   held-out CE @ 200: SFT 1.2103 (start 1.2332) | OOD 2.1123 (start 2.1218)
+[10:53:00 + 29.0m]   held-out CE @ 400: SFT 1.2084 (start 1.2332) | OOD 2.1153 (start 2.1218)
+[11:06:36 + 42.6m]   held-out CE @ 600: SFT 1.2074 (start 1.2332) | OOD 2.1141 (start 2.1218)
+[11:18:25 + 54.4m]   held-out CE @ 800: SFT 1.2017 (start 1.2332) | OOD 2.1103 (start 2.1218)
+[11:31:08 + 67.2m]   held-out CE @ 1000: SFT 1.2015 (start 1.2332) | OOD 2.1052 (start 2.1218)
+[11:44:01 + 80.0m]   held-out CE @ 1200: SFT 1.2005 (start 1.2332) | OOD 2.1059 (start 2.1218)
+[11:57:07 + 93.1m]   held-out CE @ 1400: SFT 1.1987 (start 1.2332) | OOD 2.1061 (start 2.1218)
+[12:09:22 +105.4m]   held-out CE @ 1600: SFT 1.1939 (start 1.2332) | OOD 2.1030 (start 2.1218)
+[12:22:28 +118.5m]   held-out CE @ 1800: SFT 1.1941 (start 1.2332) | OOD 2.1071 (start 2.1218)
+[12:35:30 +131.5m]   held-out CE @ 2000: SFT 1.1978 (start 1.2332) | OOD 2.1104 (start 2.1218)
+[12:48:54 +144.9m]   held-out CE @ 2200: SFT 1.2008 (start 1.2332) | OOD 2.1082 (start 2.1218)
+[12:49:55 +145.9m] TIME BUDGET: stopping at step 2201

colab2/eval/export_c2s2200.log

@@ -0,0 +1,7 @@
+merged LoRA/shared into 330 tensors
+carried 612 tensors from /content/work/train/base-sigmoid, folded 81 norms into HC
+TBL_SNAP: 7,025 of 5,859,840 non-zero rows (|row| < 0.01) exported as exact zeros
+PLE: 8 heads x ~976,640 rows x 256 = 2.00B params in 2 shards, width 2048 into hidden 2048, injected at layer 2
+EXPORT_DONE /content/work/export_c2s2200 tensors=976 shards=13 gate=sigmoid ple_layer=2
+/content/work/kit/code/wt/export_next36.py:93: UserWarning: The given NumPy array is not writable, and PyTorch does not support non-writable tensors. This means writing to this tensor will result in undefined behavior. You may want to copy the array to protect its data or make it writable before converting it to a tensor. This type of warning will be suppressed for the rest of this program. (Triggered internally at /pytorch/torch/csrc/utils/tensor_numpy.cpp:213.)
+  t = torch.from_numpy(np.ascontiguousarray(a))

colab2/eval/lw_c2s2200_held.log

@@ -0,0 +1,13 @@
+[     0s] batch (8, 512) from teacher_L59_L63_n400.npz rows 0..8 | pairs [(35, 59), (39, 63)]
+[     0s] indexer budget 262,144 -> 512 for this 512-token probe (same visible tokens)
+[     0s] c2s2200_held: 40 layers, hidden 2048, hc_count 4, attention at [3, 7, 11, 15, 19, 23, 27, 31, 35, 39]
+[     9s] streams (8, 512, 8192) | masks + rope built
+[    13s]   .. layer 0 done
+[    14s]   layer 1: n-gram memory streamed 7,813,120 rows into (7813760, 256) on cuda
+[    41s]   .. layer 8 done
+[    62s]   .. layer 16 done
+[    90s]   .. layer 24 done
+[   111s]   .. layer 32 done
+[   116s] c2s2200_held S35 <- T59: top-1 agree 20.2% | dprob 0.330 | teacher-mass 48.7% | student-mass 52.3%  (3,840 positions)
+[   121s] c2s2200_held S39 <- T63: top-1 agree 87.3% | dprob 0.082 | teacher-mass 97.2% | student-mass 97.6%  (3,840 positions)
+[   121s] LW_EVAL DONE c2s2200_held: S35<-T59 agree 20.2% mass 48.7% | S39<-T63 agree 87.3% mass 97.2%

colab2/eval/lw_c2s2200_train.log

@@ -0,0 +1,13 @@
+[     0s] batch (8, 512) from teacher_L59_L63.npz rows 0..8 | pairs [(35, 59), (39, 63)]
+[     0s] indexer budget 262,144 -> 512 for this 512-token probe (same visible tokens)
+[     0s] c2s2200_train: 40 layers, hidden 2048, hc_count 4, attention at [3, 7, 11, 15, 19, 23, 27, 31, 35, 39]
+[     1s] streams (8, 512, 8192) | masks + rope built
+[     1s]   .. layer 0 done
+[     2s]   layer 1: n-gram memory streamed 7,813,120 rows into (7813760, 256) on cuda
+[     5s]   .. layer 8 done
+[     8s]   .. layer 16 done
+[    11s]   .. layer 24 done
+[    14s]   .. layer 32 done
+[    19s] c2s2200_train S35 <- T59: top-1 agree 19.4% | dprob 0.376 | teacher-mass 50.6% | student-mass 54.0%  (3,840 positions)
+[    23s] c2s2200_train S39 <- T63: top-1 agree 82.9% | dprob 0.102 | teacher-mass 97.0% | student-mass 97.2%  (3,840 positions)
+[    23s] LW_EVAL DONE c2s2200_train: S35<-T59 agree 19.4% mass 50.6% | S39<-T63 agree 82.9% mass 97.0%

colab2/eval/lw_eval_c2s2200_held.json

@@ -0,0 +1,38 @@
+{
+ "tag": "c2s2200_held",
+ "pairs": [
+  [
+   35,
+   59
+  ],
+  [
+   39,
+   63
+  ]
+ ],
+ "cache": "/content/work/probe/teacher_L59_L63_n400.npz",
+ "row0": 0,
+ "nseq": 8,
+ "seqlen": 512,
+ "topk": 32,
+ "positions": 480,
+ "model_dir": "/content/work/export_c2s2200",
+ "results": [
+  {
+   "student": 35,
+   "teacher": 59,
+   "agree": 0.20234375,
+   "dprob": 0.33010420203208923,
+   "teacher_mass": 0.48682305216789246,
+   "student_mass": 0.5229164958000183
+  },
+  {
+   "student": 39,
+   "teacher": 63,
+   "agree": 0.87265625,
+   "dprob": 0.08171550184488297,
+   "teacher_mass": 0.9717335104942322,
+   "student_mass": 0.9761016368865967
+  }
+ ]
+}

colab2/eval/lw_eval_c2s2200_train.json

@@ -0,0 +1,38 @@
+{
+ "tag": "c2s2200_train",
+ "pairs": [
+  [
+   35,
+   59
+  ],
+  [
+   39,
+   63
+  ]
+ ],
+ "cache": "/content/work/probe/teacher_L59_L63.npz",
+ "row0": 0,
+ "nseq": 8,
+ "seqlen": 512,
+ "topk": 32,
+ "positions": 480,
+ "model_dir": "/content/work/export_c2s2200",
+ "results": [
+  {
+   "student": 35,
+   "teacher": 59,
+   "agree": 0.19375,
+   "dprob": 0.3762657344341278,
+   "teacher_mass": 0.505649745464325,
+   "student_mass": 0.5401917695999146
+  },
+  {
+   "student": 39,
+   "teacher": 63,
+   "agree": 0.82890625,
+   "dprob": 0.1018216460943222,
+   "teacher_mass": 0.969525158405304,
+   "student_mass": 0.9718742966651917
+  }
+ ]
+}