{
  "schema_version": 1,
  "configuration_role": "declarative experiment specification; CLI flags are authoritative",
  "training_host": "macm3",
  "device": "mps",
  "reference_repository": "ngxson/fly-llm-hf",
  "reference_revision": "65c677b3d566a2e9793d5f72999cdb441c6c0a9f",
  "reference_weights_sha256": "355f06c44d14e38af50e9c801f51839c37a0a56ac4ca016da2be9b3ae6f215ad",
  "dataset_path": "data/ngxson-tinystories-v1/dataset.json",
  "dataset_sha256": "c1f553263545744786a8811ee6af7ace8c452820bbb99d15140b81625e40b3a8",
  "split_stories": {
    "train": 1000,
    "validation": 100,
    "test": 100
  },
  "initialization": "all learned values from scratch; retain released graph and interface",
  "optimizer": {
    "name": "AdamW",
    "betas": [
      0.9,
      0.999
    ],
    "epsilon": 1e-08,
    "weight_decay": 0.01,
    "gradient_clip_norm": 1.0,
    "phases": [
      {
        "epochs": 30,
        "input_neuron_lr": 0.001,
        "readout_lr": 0.0001
      },
      {
        "epochs": 14,
        "input_neuron_lr": 0.0003,
        "readout_lr": 3e-05
      }
    ],
    "schedule": "cosine to zero separately per phase; preserve Adam moments across phases"
  },
  "checkpoint_selection": {
    "primary": "minimum token-weighted validation cross-entropy",
    "secondary": "maximum validation top-1 accuracy; retain actual weights and selection receipt"
  },
  "claim_scope": "User-selected rank32 decoder comparison against preserved rank64; quality not yet measured",
  "environment": {
    "PYTORCH_ENABLE_MPS_FALLBACK": "0",
    "OMP_NUM_THREADS": "4",
    "VECLIB_MAXIMUM_THREADS": "4"
  },
  "experiment_id": "H32rank32fixed",
  "branch": "exp/encoder32-readout32-fixed",
  "trainer": "scripts/train_connectorch.py",
  "connectorch_revision": "4bbfb645099aeb85bdbf850e1a87cc094769af87",
  "groups_path": "data/connectorch-groups-v1/groups.npz",
  "groups_sha256": "a07d311cd04126b5e54ea9cb3e8945470af9049bcb8499fb406e3c38d185dd7f",
  "trainable_parameters": 2343125,
  "training": {
    "device": "mps",
    "epochs": 30,
    "second_epochs": 14,
    "batch_size": 8,
    "chunk_size": 32,
    "seed": 42,
    "eval_interval_updates": 100,
    "d_embed": 32,
    "plasticity": "fixed",
    "readout_rank": 32,
    "history_length": 8,
    "skip_final_test": true
  },
  "edge_gain_parameters": 0,
  "edge_multiplier_bounds": [
    1.0,
    1.0
  ],
  "neuron_gains": "original unconstrained gain and rec_gain",
  "test_policy": "deferred until comparison and final protocol are fixed",
  "output_relative": "results/connectorch-rank32-v1/arms/H32rank32fixed",
  "readout_parameters": 1613344,
  "encoder_parameters": 482816,
  "comparison_baseline": "G32rank64fixed; compare common training target budgets and both independent validation selectors",
  "readout_architecture": {
    "input_neurons": 49393,
    "rank": 32,
    "vocabulary": 1024,
    "bias": false,
    "intermediate_activation": null
  },
  "initialization_note": "Same from-scratch trainer initialization and seed42; preserve encoder and neuron initialization. Rank32 changes readout shapes and initial logit variance relative to rank64; verify rank32-specific gradients and optimizer updates before training.",
  "execution_status": "selected by user; separate rank32 native CPU/MPS preflight and optimizer smoke required",
  "prerequisites": [
    "Confirm idle macm3 and preserved G accepted stop with both winners/latest before H launch",
    "Run separate rank32 CPU/MPS parity and eight-update optimizer smoke on macm3 before full H training"
  ]
}