btx-e003 R3 ledger

mini-beatrix-1/reports/arms/btx_e003_R3_ledger.json

@@ -0,0 +1,338 @@
+{
+ "rung": "btx-e003-R3",
+ "started": "2026-09-06 08:22",
+ "teacher_gate": {
+  "instructed_strict": true,
+  "uninstructed_chain": true,
+  "final_predicate_any": true
+ },
+ "restricted_targets": false,
+ "lm_clean": {
+  "nextbyte": 0.5679,
+  "bpb": 2.1786
+ },
+ "floors_from_R0": [
+  {
+   "arm": "bare@88508",
+   "exam150": 0.0,
+   "exam30": 0.0
+  },
+  {
+   "arm": "bare@58664",
+   "exam150": 0.0,
+   "exam30": 0.0
+  },
+  {
+   "arm": "d5-XL",
+   "step": 88508,
+   "params": 3185008,
+   "spec": {
+    "n_slots": 16,
+    "K": 64,
+    "D": 4,
+    "hidden": 178
+   },
+   "exam150": 0.4,
+   "exam30": 1.0,
+   "exam150_v0_wordform_shift": 0.20666666666666667
+  },
+  {
+   "arm": "d5-XL-wide",
+   "step": 88508,
+   "params": 5275664,
+   "spec": {
+    "n_slots": 32,
+    "K": 64,
+    "D": 4,
+    "hidden": 256
+   },
+   "exam150": 0.47333333333333333,
+   "exam30": 1.0,
+   "exam150_v0_wordform_shift": 0.23333333333333334
+  },
+  {
+   "arm": "loopctrl-d5-s0",
+   "step": 88508,
+   "params": 5275664,
+   "spec": {
+    "n_slots": 32,
+    "K": 64,
+    "D": 4,
+    "hidden": 256
+   },
+   "exam150": 0.1,
+   "exam30": 0.8333333333333334,
+   "exam150_v0_wordform_shift": 0.09333333333333334
+  },
+  {
+   "arm": "d5ctrl-s0",
+   "step": 58664,
+   "params": 3185008,
+   "spec": {
+    "n_slots": 16,
+    "K": 64,
+    "D": 4,
+    "hidden": 178
+   },
+   "exam150": 0.44,
+   "exam30": 0.8,
+   "exam150_v0_wordform_shift": 0.24666666666666667
+  },
+  {
+   "arm": "d5ctrl-s1",
+   "step": 58664,
+   "params": 3185008,
+   "spec": {
+    "n_slots": 16,
+    "K": 64,
+    "D": 4,
+    "hidden": 178
+   },
+   "exam150": 0.43333333333333335,
+   "exam30": 0.6333333333333333,
+   "exam150_v0_wordform_shift": 0.30666666666666664
+  }
+ ],
+ "gauge_only": [
+  {
+   "arm": "d5-XL",
+   "step": 88508,
+   "params": 3185008,
+   "spec": {
+    "n_slots": 16,
+    "K": 64,
+    "D": 4,
+    "hidden": 178
+   },
+   "meta": {
+    "step": 88508,
+    "name": "d5-XL",
+    "group": "day1",
+    "kind": "behavior",
+    "behavior": "d5-XL",
+    "sites": 16,
+    "params": 3185008,
+    "base_model_id": "alephllm/mini-beatrix-1@step88508",
+    "verified": true,
+    "template": {
+     "frame": "qa",
+     "stops": [
+      "\n\n",
+      "\nUser:",
+      "\nBeatrix:",
+      "User:",
+      "Beatrix:"
+     ],
+     "chat_ok": true,
+     "turn_end": "newline pair",
+     "source": "overnight_arms.row(): single-turn QA frame, reply + b'\\n\\n'"
+    },
+    "decode": {
+     "temperature": 0.0,
+     "top_p": 1.0,
+     "why": "task arms collapse to turn-end off-distribution and degenerate into punctuation under sampling (measured 2026-08-19) \u2014 greedy is their honest setting"
+    },
+    "examples": [
+     "If someone is glent, then they are sorb. If someone is sorb, then they are trell. Wren is glent. What follows?",
+     "Rules: If someone is vint, then they are kresp. If someone is kresp, then they are dorn. Fact: Nils is vint. Conclusion?"
+    ]
+   },
+   "exam150": 0.94,
+   "exam150_n_answers": 5.06,
+   "exam30": 1.0,
+   "exam30_n_answers": 5.0,
+   "exam150_v1_novel_aliens": 0.4,
+   "exam150_v1_novel_aliens_n_answers": 5.22,
+   "exam150_v0_wordform_shift": 0.20666666666666667,
+   "exam150_v0_wordform_shift_n_answers": 5.133,
+   "interference": {
+    "nextbyte_on": 0.5822,
+    "bpb_on": 2.5533,
+    "bpb_clean": 2.1786,
+    "bpb_delta": 0.3747,
+    "nextbyte_clean": 0.5679
+   }
+  },
+  {
+   "arm": "d5-XL-wide",
+   "step": 88508,
+   "params": 5275664,
+   "spec": {
+    "n_slots": 32,
+    "K": 64,
+    "D": 4,
+    "hidden": 256
+   },
+   "meta": {
+    "step": 88508,
+    "name": "d5-XL-wide",
+    "group": "day1",
+    "kind": "behavior",
+    "behavior": "d5-XL-wide",
+    "sites": 16,
+    "params": 5275664,
+    "base_model_id": "alephllm/mini-beatrix-1@step88508",
+    "verified": true,
+    "template": {
+     "frame": "qa",
+     "stops": [
+      "\n\n",
+      "\nUser:",
+      "\nBeatrix:",
+      "User:",
+      "Beatrix:"
+     ],
+     "chat_ok": true,
+     "turn_end": "newline pair",
+     "source": "overnight_arms.row(): single-turn QA frame, reply + b'\\n\\n'"
+    },
+    "decode": {
+     "temperature": 0.0,
+     "top_p": 1.0,
+     "why": "task arms collapse to turn-end off-distribution and degenerate into punctuation under sampling (measured 2026-08-19) \u2014 greedy is their honest setting"
+    },
+    "examples": [
+     "If someone is glent, then they are sorb. If someone is sorb, then they are trell. Wren is glent. What follows?",
+     "Rules: If someone is vint, then they are kresp. If someone is kresp, then they are dorn. Fact: Nils is vint. Conclusion?"
+    ]
+   },
+   "exam150": 1.0,
+   "exam150_n_answers": 5.0,
+   "exam30": 1.0,
+   "exam30_n_answers": 5.0,
+   "exam150_v1_novel_aliens": 0.47333333333333333,
+   "exam150_v1_novel_aliens_n_answers": 5.033,
+   "exam150_v0_wordform_shift": 0.23333333333333334,
+   "exam150_v0_wordform_shift_n_answers": 4.78,
+   "interference": {
+    "nextbyte_on": 0.5828,
+    "bpb_on": 2.5352,
+    "bpb_clean": 2.1786,
+    "bpb_delta": 0.3566,
+    "nextbyte_clean": 0.5679
+   }
+  }
+ ],
+ "arms": [
+  {
+   "arm": "mb-T1-s0",
+   "torch_seed": 719195110,
+   "restricted": false,
+   "descent_gate": {
+    "passed": true,
+    "first": 0.402341,
+    "last": 6.5e-05,
+    "rel_drop": 0.9998
+   },
+   "train_s": 3716,
+   "curve": [
+    [
+     250,
+     0.14236
+    ],
+    [
+     500,
+     0.03115
+    ],
+    [
+     750,
+     0.02579
+    ],
+    [
+     1000,
+     0.02036
+    ],
+    [
+     1250,
+     0.01346
+    ],
+    [
+     1500,
+     0.01316
+    ],
+    [
+     1750,
+     0.00972
+    ],
+    [
+     2000,
+     0.00956
+    ],
+    [
+     2250,
+     0.00994
+    ],
+    [
+     2500,
+     0.0082
+    ],
+    [
+     2750,
+     0.00711
+    ],
+    [
+     3000,
+     0.01392
+    ],
+    [
+     3250,
+     0.01091
+    ],
+    [
+     3500,
+     0.0075
+    ],
+    [
+     3750,
+     0.0075
+    ],
+    [
+     4000,
+     0.01151
+    ]
+   ],
+   "target_stats": {
+    "positions": 2714921,
+    "token_start_fraction": 0.4125,
+    "entropy_at_starts": 0.0903,
+    "entropy_mid_token": 0.1743,
+    "argmax_ne_gold_fraction": 0.107,
+    "fallback_fraction": 0.0011,
+    "restricted_fraction": 0.0
+   },
+   "exam150": 0.0,
+   "exam150_n_answers": 0.0,
+   "exam30": 0.0,
+   "exam30_n_answers": 0.0,
+   "exam150_v1_novel_aliens": 0.0,
+   "exam150_v1_novel_aliens_n_answers": 0.0,
+   "exam150_v0_wordform_shift": 0.0,
+   "exam150_v0_wordform_shift_n_answers": 0.0,
+   "interference": {
+    "nextbyte_on": 0.5789,
+    "bpb_on": 2.2834,
+    "bpb_clean": 2.1786,
+    "bpb_delta": 0.1048,
+    "nextbyte_clean": 0.5679
+   },
+   "bitexact_off": true,
+   "anchor": "E:\\mirel\\data\\btx_e003\\arms\\btx-e003-R3-mb-T1-s0@step88508.anchor.pt",
+   "alloy_ledger": {
+    "mb": {
+     "d5": {
+      "n": 4000,
+      "mean": 0.036229
+     }
+    }
+   },
+   "peak_vram_gb": 8.23
+  }
+ ],
+ "_env": {
+  "torch": "2.11.0+cu128",
+  "gpu": "NVIDIA GeForce RTX 4090",
+  "steps": 4000,
+  "batch": 8,
+  "spec": "WIDE n32 K64 D4 h256"
+ }
+}