btx-e003 R0 instruments: r0_done.json

btx_e003/instruments/r0_done.json

@@ -0,0 +1,225 @@
+{
+ "started": "2026-09-05 08:26",
+ "assembled": "2026-09-05 10:23",
+ "stages": {
+  "S0": {
+   "from": "r0.log",
+   "step1_delta": 0.0,
+   "hand": 0.3833617,
+   "lib": 0.3833617,
+   "broken_closure_failed": true,
+   "oracle_gate": {
+    "passed": true,
+    "first": 0.397058,
+    "last": 0.000839,
+    "rel_drop": 0.9979
+   },
+   "no_adamw_grep": "PASSED"
+  },
+  "S1": {
+   "n150": 150,
+   "rows": [
+    {
+     "arm": "bare@88508",
+     "exam150": 0.0,
+     "exam30": 0.0
+    },
+    {
+     "arm": "bare@58664",
+     "exam150": 0.0,
+     "exam30": 0.0
+    },
+    {
+     "arm": "d5-XL-wide",
+     "error": "Error(s) in loading state_dict for RelayPatchwork:\n\tsize mismatch for proj.weight: copying a param with shape torch.Size([128, 768]) from checkpoint, the shape in current model is torch.Size([64, 768]"
+    },
+    {
+     "arm": "d5-XL-wide",
+     "error": "Error(s) in loading state_dict for RelayPatchwork:\n\tsize mismatch for proj.weight: copying a param with shape torch.Size([128, 768]) from checkpoint, the shape in current model is torch.Size([64, 768]"
+    },
+    {
+     "arm": "loopctrl-d5-s0",
+     "error": "Error(s) in loading state_dict for RelayPatchwork:\n\tsize mismatch for proj.weight: copying a param with shape torch.Size([128, 768]) from checkpoint, the shape in current model is torch.Size([64, 768]"
+    },
+    {
+     "arm": "d5ctrl-s0",
+     "step": 58664,
+     "params": 3185008,
+     "exam150": 0.24666666666666667,
+     "exam30": 0.8
+    },
+    {
+     "arm": "d5ctrl-s1",
+     "step": 58664,
+     "params": 3185008,
+     "exam150": 0.30666666666666664,
+     "exam30": 0.6333333333333333
+    }
+   ],
+   "peak_vram_gb": 1.86
+  },
+  "S2": {
+   "teacher_d5_exam150": 0.0,
+   "teacher_d5_exam30": 0.0,
+   "teacher_d5_held200": 0.0,
+   "gate_teacher_d5_ge_.90": false,
+   "d5": {
+    "n": 24000,
+    "bytes_ok": 24000,
+    "max_T": 35,
+    "mean_T": 35.0,
+    "path": "E:\\mirel\\data\\btx_e003\\bank_v3_d5_qwen25_3b.npz",
+    "crc32": "a4cff722"
+   },
+   "sub3": {
+    "n": 20050,
+    "bytes_ok": 20050,
+    "max_T": 3,
+    "mean_T": 2.7761596009975062,
+    "path": "E:\\mirel\\data\\btx_e003\\bank_v3_sub3_qwen25_3b.npz",
+    "crc32": "5536801c"
+   },
+   "peak_vram_gb": 6.7
+  },
+  "S3": {
+   "d5": {
+    "rows": 4000,
+    "positions": 339326,
+    "mid_token_positions": 199326,
+    "mid_token_fraction": 0.5874174098064988,
+    "jsd_mid_median": 0.6931471672045528,
+    "jsd_mid_p95": 0.6931471805599454,
+    "jsd_lt_.005_fraction": 0.009913408185585422,
+    "gold_byte_argmax_rate": 0.8921803811084326,
+    "dropped_mass_median": 1.1812154835758568e-07,
+    "dropped_mass_p95": 0.7010381780887871,
+    "a0_fixtures": null,
+    "sparse_conditional": "E:\\mirel\\data\\btx_e003\\byte_conditional_top16_d5.npz"
+   },
+   "sub3": {
+    "rows": 4050,
+    "positions": 11244,
+    "mid_token_positions": 0,
+    "mid_token_fraction": 0.0,
+    "jsd_mid_median": 0.0,
+    "jsd_mid_p95": 0.0,
+    "jsd_lt_.005_fraction": 1.0,
+    "gold_byte_argmax_rate": 0.9874599786552828,
+    "dropped_mass_median": 1.1321480164005255e-19,
+    "dropped_mass_p95": 8.258072338711938e-14,
+    "a0_fixtures": "140/140",
+    "sparse_conditional": "E:\\mirel\\data\\btx_e003\\byte_conditional_top16_sub3.npz"
+   },
+   "gate_note": "The multi-byte form is exact within a token from the top-64; off-gold re-entry (a shorter token ending inside the gold token) is DISCLOSED as dropped mass, not beamed (the Vieira beam needs extra teacher forwards; owed). At token boundaries the gold-path next byte is the next position's first byte. The JSD gate is scored at mid-token positions only (k >= 1); where mid_token_fraction ~ 0 (sub3 digits: Qwen2.5 has no ASCII multi-digit tokens) channel 1 is LABEL-CLASS.",
+   "gate_pass_d5": false
+  },
+  "S4": {
+   "codex_C": {
+    "texts": 2296,
+    "walk_ok": 2296,
+    "tokens": 77436,
+    "bytes": 253720,
+    "bytes_per_token": 3.2765122165401106,
+    "mid_token_fraction": 0.694797414472647,
+    "word_boundaries": 55464,
+    "unaligned_word_boundary_fraction": 0.8898745131977499,
+    "spacelike_vs_teacher_boundaries": {
+     "precision": 0.11012548680225011,
+     "recall": 0.07887804122113745,
+     "f1": 0.09191873589164787
+    },
+    "chunk_cells": {
+     "1x1": 0,
+     "nm": 6108
+    }
+   },
+   "d5_replies": {
+    "texts": 5000,
+    "walk_ok": 5000,
+    "tokens": 175000,
+    "bytes": 424223,
+    "bytes_per_token": 2.4241314285714286,
+    "mid_token_fraction": 0.5874811125280808,
+    "word_boundaries": 100000,
+    "unaligned_word_boundary_fraction": 0.95,
+    "spacelike_vs_teacher_boundaries": {
+     "precision": 0.05,
+     "recall": 0.02857142857142857,
+     "f1": 0.03636363636363636
+    },
+    "chunk_cells": {
+     "1x1": 0,
+     "nm": 5000
+    }
+   },
+   "d5_questions": {
+    "texts": 5000,
+    "walk_ok": 5000,
+    "tokens": 345000,
+    "bytes": 1152493,
+    "bytes_per_token": 3.340559420289855,
+    "mid_token_fraction": 0.7006489410347828,
+    "word_boundaries": 225000,
+    "unaligned_word_boundary_fraction": 0.9777777777777777,
+    "spacelike_vs_teacher_boundaries": {
+     "precision": 0.022222222222222223,
+     "recall": 0.014492753623188406,
+     "f1": 0.01754385964912281
+    },
+    "chunk_cells": {
+     "1x1": 0,
+     "nm": 5000
+    }
+   },
+   "sub3_replies": {
+    "texts": 5000,
+    "walk_ok": 5000,
+    "tokens": 13919,
+    "bytes": 13919,
+    "bytes_per_token": 1.0,
+    "mid_token_fraction": 0.0,
+    "word_boundaries": 5000,
+    "unaligned_word_boundary_fraction": 0.0,
+    "spacelike_vs_teacher_boundaries": {
+     "precision": 1.0,
+     "recall": 0.3592212084201451,
+     "f1": 0.5285691632750145
+    },
+    "chunk_cells": {
+     "1x1": 106,
+     "nm": 4894
+    }
+   },
+   "note": "co-boundaries = word boundaries that are also teacher token boundaries; cells between consecutive co-boundaries; the spacelike rule (space -> non-space) vs the teacher's own boundaries is the zero-parameter control (S-701/S-702). The cond_entropy endpoint comparison (bytelex matrix tools) is owed in r0b."
+  },
+  "S5": {
+   "pairwise_cos": {
+    "byte_kl_T1|byte_kl_T4_smoothed": -0.11049807071685791,
+    "byte_kl_T1|chunk_1x1": 0.9999750852584839,
+    "byte_kl_T1|ce": 1.0,
+    "byte_kl_T4_smoothed|chunk_1x1": -0.1121935173869133,
+    "byte_kl_T4_smoothed|ce": -0.11050013452768326,
+    "ce|chunk_1x1": 0.9999749660491943
+   },
+   "novelty": {
+    "byte_kl_T1": 0.0,
+    "byte_kl_T4_smoothed": 0.8878064826130867,
+    "chunk_1x1": 2.491474151611328e-05,
+    "ce": 0.0
+   },
+   "refuse_lt_.05": [
+    "byte_kl_T1",
+    "chunk_1x1",
+    "ce"
+   ],
+   "merge_gt_.8": [
+    "byte_kl_T1|chunk_1x1",
+    "byte_kl_T1|ce",
+    "ce|chunk_1x1"
+   ],
+   "peak_vram_gb": 4.57,
+   "note": "grads on 4 batches x 4 rows of oracle d5 through the fresh WIDE arm at all 16 blocks (autograd.grad per component on the shared forward); byte_kl_T1 / chunk_1x1 / ce are the same objective on ONE-HOT oracles and are expected collinear here \u2014 the gate's bite is on real targets (R3)."
+  }
+ },
+ "finished": "2026-09-05 10:23"
+}