btx-e003 2s ladder C5/C4/C6 ledger

mini-beatrix-2s/reports/arms/btx_e003_C456_ledger.json

@@ -34385,7 +34385,103 @@
     "peak_vram_gb": 5.26,
     "finished": "2026-09-20 01:52"
    },
-   "error": "tools\\btx_2s_ladder_e_i.py\", line 549, in <module>\n    train_frame_member(FR, f\"btx-e003-2s-EI-frame-{S}\", {\"chain\": d5_abst, \"stop\": chat_abst, \"neutral\": fw_abst}, DOM)\n  File \"E:\\mirel\\claude\\history\\dockets\\2026-09-05_bytelex_plan\\tools\\btx_2s_ladder_e_i.py\", line 342, in train_frame_member\n    la = LAMBDA * kl_abstain(m, handle, rows[int(rng.integers(0, len(rows)))]); (la / 2).backward(); ta += float(la) / 2\n                                                                                ^^^^^^^^^^^^^^^^^^^\n  File \"E:\\mirel\\claude\\.venv\\Lib\\site-packages\\torch\\_tensor.py\", line 631, in backward\n    torch.autograd.backward(\n  File \"E:\\mirel\\claude\\.venv\\Lib\\site-packages\\torch\\autograd\\__init__.py\", line 381, in backward\n    _engine_run_backward(\n  File \"E:\\mirel\\claude\\.venv\\Lib\\site-packages\\torch\\autograd\\graph.py\", line 869, in _engine_run_backward\n    return Variable._execution_engine.run_backward(  # Calls into the C++ engine to run the backward pass\n           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\ntorch.AcceleratorError: CUDA error: unknown error\nSearch for `cudaErrorUnknown' in https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html for more information.\nCUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.\nFor debugging consider passing CUDA_LAUNCH_BLOCKING=1\nCompile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.\n\n"
+   "error_previous_launch": "tools\\btx_2s_ladder_e_i.py\", line 549, in <module>\n    train_frame_member(FR, f\"btx-e003-2s-EI-frame-{S}\", {\"chain\": d5_abst, \"stop\": chat_abst, \"neutral\": fw_abst}, DOM)\n  File \"E:\\mirel\\claude\\history\\dockets\\2026-09-05_bytelex_plan\\tools\\btx_2s_ladder_e_i.py\", line 342, in train_frame_member\n    la = LAMBDA * kl_abstain(m, handle, rows[int(rng.integers(0, len(rows)))]); (la / 2).backward(); ta += float(la) / 2\n                                                                                ^^^^^^^^^^^^^^^^^^^\n  File \"E:\\mirel\\claude\\.venv\\Lib\\site-packages\\torch\\_tensor.py\", line 631, in backward\n    torch.autograd.backward(\n  File \"E:\\mirel\\claude\\.venv\\Lib\\site-packages\\torch\\autograd\\__init__.py\", line 381, in backward\n    _engine_run_backward(\n  File \"E:\\mirel\\claude\\.venv\\Lib\\site-packages\\torch\\autograd\\graph.py\", line 869, in _engine_run_backward\n    return Variable._execution_engine.run_backward(  # Calls into the C++ engine to run the backward pass\n           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\ntorch.AcceleratorError: CUDA error: unknown error\nSearch for `cudaErrorUnknown' in https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html for more information.\nCUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.\nFor debugging consider passing CUDA_LAUNCH_BLOCKING=1\nCompile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.\n\n",
+   "frame-sel-sL1": {
+    "arm": "frame-sel-sL1",
+    "cell": "E-I",
+    "torch_seed": 1185851399,
+    "seed_key": "btx-e003-2s-EI-frame-sL1",
+    "lambda": 1.0,
+    "steps": 6000,
+    "train_mode": {
+     "batch": 2,
+     "accum": 4,
+     "hub_ckpt": 0,
+     "bf16_replication": false
+    },
+    "description": "frame-sel: the Cell 5 alloy (L_id + MSE to bert L4-first / t5-base L6-last) at block 9 on codex lines + abstention KL on d5 rows, chat rows and fineweb slices",
+    "recipe": "pure Adam wd0 lr1e-3 6000 steps, 16 codex lines per step as 4 chunks of 4 (site-weighted) + 2 abstention rows per step (lambda 1.0); WIDE-1024 bias-zeroed",
+    "frame": "text",
+    "template": "qa newline pair: CHAT_HEADER + 'User: ' + q + '\\n' + 'Beatrix: ' + reply + '\\n\\n' (amoe_bridge.py:30-34)",
+    "stops": [
+     "\nUser:",
+     "\nBeatrix:",
+     "User:",
+     "Beatrix:",
+     "\n\n"
+    ],
+    "turn_end": "newline pair [10,10]",
+    "trunk_sha256_before": "95a4606ff34dbcc0d907540303796bf080010ca8708c013bd18ac9b34107ecf5",
+    "train_s": 5514,
+    "s_per_step": 0.919,
+    "curve": [
+     [
+      500,
+      2.9597,
+      0.00103
+     ],
+     [
+      1000,
+      2.48944,
+      0.00082
+     ],
+     [
+      1500,
+      2.30284,
+      0.00222
+     ],
+     [
+      2000,
+      2.09938,
+      0.00236
+     ],
+     [
+      2500,
+      2.07689,
+      0.00069
+     ],
+     [
+      3000,
+      1.95861,
+      0.00436
+     ],
+     [
+      3500,
+      2.02277,
+      0.00101
+     ],
+     [
+      4000,
+      1.98346,
+      0.00228
+     ],
+     [
+      4500,
+      1.9928,
+      0.00214
+     ],
+     [
+      5000,
+      1.82477,
+      0.00394
+     ],
+     [
+      5500,
+      1.82199,
+      0.00252
+     ],
+     [
+      6000,
+      1.84793,
+      0.004
+     ]
+    ],
+    "peak_vram_train_gb": 4.07,
+    "max_abs_dw_from_init": 3.6740598678588867,
+    "anchor": "E:\\mirel\\data\\btx_e003\\arms_2s\\btx-e003-2s-frame-sel-sL1.b2a4@step61422.anchor.pt",
+    "trainable_params": 8558100,
+    "provisional": true
+   }
   },
   "E-D0": {
    "started": "2026-09-10 11:28",