{
  "tag": "mdlm",
  "text": "The secret word is apple.",
  "tokens": [
    "The",
    " secret",
    " word",
    " is",
    " apple",
    "."
  ],
  "input_ids": [
    464,
    3200,
    1573,
    318,
    50257,
    13
  ],
  "input_display": [
    "The",
    " secret",
    " word",
    " is",
    "[MASK]",
    "."
  ],
  "focus_position": 2,
  "focus_token": " word",
  "focus_id": 1573,
  "target_id": 17180,
  "target_token": " apple",
  "prediction_position": 4,
  "layers": 12,
  "embedding_shape": [
    50258,
    768
  ],
  "embedding_samples": [
    {
      "coordinate": 562,
      "before": -0.13285495340824127,
      "gradient": 1.9260504245758057,
      "after": -0.15211546421051025
    },
    {
      "coordinate": 4,
      "before": -0.20384244620800018,
      "gradient": 1.7183243036270142,
      "after": -0.22102569043636322
    },
    {
      "coordinate": 605,
      "before": 0.21036402881145477,
      "gradient": -1.696223497390747,
      "after": 0.22732625901699066
    },
    {
      "coordinate": 579,
      "before": 0.15997397899627686,
      "gradient": 1.6825306415557861,
      "after": 0.14314867556095123
    }
  ],
  "embedding_delta_norm": 0.14298370480537415,
  "parameters": [
    {
      "name": "backbone.blocks.11.mlp.2.weight",
      "shape": [
        768,
        3072
      ],
      "gradient_norm": 2.124068021774292,
      "changed_coefficients": 2347560,
      "samples": [
        {
          "coordinate": [
            423,
            3019
          ],
          "before": -0.1653643548488617,
          "gradient": 0.04526538401842117,
          "after": -0.1658170074224472
        },
        {
          "coordinate": [
            759,
            3019
          ],
          "before": -0.1487920731306076,
          "gradient": -0.045112524181604385,
          "after": -0.14834094047546387
        },
        {
          "coordinate": [
            608,
            3019
          ],
          "before": 0.13283567130565643,
          "gradient": 0.04351983964443207,
          "after": 0.13240046799182892
        },
        {
          "coordinate": [
            218,
            3019
          ],
          "before": 0.21377378702163696,
          "gradient": -0.04251878336071968,
          "after": 0.2141989767551422
        }
      ]
    }
  ],
  "before": {
    "ce": 16.44001579284668,
    "loss": 16.44001579284668,
    "target_probability": 7.247562194834245e-08,
    "top5": [
      {
        "id": 337,
        "text": " M",
        "p": 0.16244545578956604
      },
      {
        "id": 440,
        "text": " O",
        "p": 0.061374448239803314
      },
      {
        "id": 471,
        "text": " U",
        "p": 0.06071433797478676
      },
      {
        "id": 370,
        "text": " W",
        "p": 0.058688584715127945
      },
      {
        "id": 317,
        "text": " A",
        "p": 0.05736061930656433
      }
    ],
    "layers": {
      "0": {
        "rms": 1.8727329969406128,
        "first4": [
          -0.004387792199850082,
          0.9745478630065918,
          -0.4018712639808655,
          0.04620622470974922
        ]
      },
      "1": {
        "rms": 3.538815498352051,
        "first4": [
          0.040148720145225525,
          -0.19287192821502686,
          -1.1463944911956787,
          -0.08104605972766876
        ]
      },
      "2": {
        "rms": 4.01917028427124,
        "first4": [
          -0.1000509262084961,
          -0.9116992950439453,
          -2.066596508026123,
          0.004194140434265137
        ]
      },
      "3": {
        "rms": 4.750522136688232,
        "first4": [
          -1.5258450508117676,
          -0.8833329081535339,
          -2.4149680137634277,
          0.44090110063552856
        ]
      },
      "4": {
        "rms": 5.683810710906982,
        "first4": [
          -3.8453621864318848,
          -0.37177032232284546,
          -3.136157989501953,
          0.33474448323249817
        ]
      },
      "5": {
        "rms": 7.43612813949585,
        "first4": [
          -6.27165412902832,
          -0.5564625263214111,
          -3.207261323928833,
          2.1975669860839844
        ]
      },
      "6": {
        "rms": 8.560951232910156,
        "first4": [
          -7.8316473960876465,
          -0.45348769426345825,
          -2.650439739227295,
          2.661762237548828
        ]
      },
      "7": {
        "rms": 12.130353927612305,
        "first4": [
          -15.710151672363281,
          -0.30887019634246826,
          -1.5547521114349365,
          5.636063575744629
        ]
      },
      "8": {
        "rms": 14.071083068847656,
        "first4": [
          -27.264476776123047,
          0.35869672894477844,
          1.3629941940307617,
          13.915426254272461
        ]
      },
      "9": {
        "rms": 20.926231384277344,
        "first4": [
          -53.84423065185547,
          -1.0324866771697998,
          7.580178737640381,
          21.473478317260742
        ]
      },
      "10": {
        "rms": 33.76814270019531,
        "first4": [
          -79.63041687011719,
          4.145811557769775,
          6.357814788818359,
          34.006103515625
        ]
      },
      "11": {
        "rms": 57.13154220581055,
        "first4": [
          -157.09715270996094,
          7.022965431213379,
          4.674568176269531,
          34.51872253417969
        ]
      }
    },
    "routes": {}
  },
  "after": {
    "ce": 14.581281661987305,
    "loss": 14.581281661987305,
    "target_probability": 4.6497504513354215e-07,
    "top5": [
      {
        "id": 337,
        "text": " M",
        "p": 0.12795162200927734
      },
      {
        "id": 471,
        "text": " U",
        "p": 0.07474784553050995
      },
      {
        "id": 440,
        "text": " O",
        "p": 0.07308562844991684
      },
      {
        "id": 509,
        "text": " K",
        "p": 0.045282673090696335
      },
      {
        "id": 370,
        "text": " W",
        "p": 0.039969705045223236
      }
    ],
    "layers": {
      "0": {
        "rms": 1.8807697296142578,
        "first4": [
          -0.006700286641716957,
          1.0080662965774536,
          -0.38880401849746704,
          0.04734569787979126
        ]
      },
      "1": {
        "rms": 3.546527624130249,
        "first4": [
          0.07646682858467102,
          -0.20682448148727417,
          -1.078771710395813,
          -0.09835562855005264
        ]
      },
      "2": {
        "rms": 4.043206214904785,
        "first4": [
          -0.0728406086564064,
          -0.9883164763450623,
          -2.051114320755005,
          0.013192795217037201
        ]
      },
      "3": {
        "rms": 4.786197662353516,
        "first4": [
          -1.4705196619033813,
          -0.948164165019989,
          -2.3494296073913574,
          0.41194918751716614
        ]
      },
      "4": {
        "rms": 5.73084831237793,
        "first4": [
          -3.527615785598755,
          -0.4912666380405426,
          -3.1585288047790527,
          0.2234712541103363
        ]
      },
      "5": {
        "rms": 7.425413608551025,
        "first4": [
          -5.949721813201904,
          -0.5802489519119263,
          -3.2652347087860107,
          2.0647153854370117
        ]
      },
      "6": {
        "rms": 8.632805824279785,
        "first4": [
          -7.7045464515686035,
          -0.5259144306182861,
          -2.8886189460754395,
          1.8891953229904175
        ]
      },
      "7": {
        "rms": 11.940580368041992,
        "first4": [
          -14.71815013885498,
          -0.39752498269081116,
          -1.8757121562957764,
          2.9609322547912598
        ]
      },
      "8": {
        "rms": 14.488143920898438,
        "first4": [
          -29.166221618652344,
          0.07060573995113373,
          1.5409332513809204,
          9.604376792907715
        ]
      },
      "9": {
        "rms": 20.890247344970703,
        "first4": [
          -52.69833755493164,
          -2.2417914867401123,
          7.242060661315918,
          11.678590774536133
        ]
      },
      "10": {
        "rms": 33.83192443847656,
        "first4": [
          -83.37738037109375,
          2.3200149536132812,
          4.879008769989014,
          21.565401077270508
        ]
      },
      "11": {
        "rms": 55.83977127075195,
        "first4": [
          -161.6463165283203,
          5.145533561706543,
          2.3454747200012207,
          15.888168334960938
        ]
      }
    },
    "routes": {}
  },
  "vocab_preview": [
    {
      "id": 17180,
      "text": " apple",
      "before": 7.247562194834245e-08,
      "after": 4.6497504513354215e-07
    },
    {
      "id": 337,
      "text": " M",
      "before": 0.16244545578956604,
      "after": 0.12795162200927734
    },
    {
      "id": 440,
      "text": " O",
      "before": 0.061374448239803314,
      "after": 0.07308562844991684
    },
    {
      "id": 471,
      "text": " U",
      "before": 0.06071433797478676,
      "after": 0.07474784553050995
    },
    {
      "id": 370,
      "text": " W",
      "before": 0.058688584715127945,
      "after": 0.039969705045223236
    },
    {
      "id": 317,
      "text": " A",
      "before": 0.05736061930656433,
      "after": 0.026540357619524002
    },
    {
      "id": 509,
      "text": " K",
      "before": 0.05349342152476311,
      "after": 0.045282673090696335
    }
  ],
  "learning_rate": 0.01,
  "scope": "ONE example, ONE local SGD step. Effective input embedding row via additive delta (both tied output/input uses for Qwen and Granite), last FFN output matrix; Granite last router also. All other base weights frozen. This is NOT the original full pretraining run.",
  "compute_dtype": "torch.float32",
  "checkpoint_modified": false,
  "ids_changed": false,
  "optimizer": "single SGD step with backtracking over fixed candidate step sizes; not a training recipe",
  "loss_definition": "single masked-position cross entropy"
}