{
  "control": "ngram5_freq_gap",
  "num_params": 1278819840,
  "num_steps": 2000,
  "peak_vram_mb": 51063.46533203125,
  "run_contract": {
    "control": "ngram5_freq_gap",
    "dataset": {
      "alpha": 0.0,
      "block_alignment_version": 2,
      "block_len": 7,
      "context_len": 5,
      "doc_len": 2048,
      "frequency_definition": "exact_train_epoch_context_count",
      "frequency_key_type": "exact_context",
      "frequency_source_split": "train",
      "hash_bucket_occupancy_diagnostic": false,
      "loader_row_stride": 2044,
      "loader_rows_train": 170264,
      "loader_rows_val": 1675858,
      "n_nonempty_buckets": 43039820,
      "order": 5,
      "resampling": "none",
      "schema_version": 1,
      "sep_token": 8191,
      "source": "code/make_ngram_blocks.py (sliding window, one copy per event)",
      "stream_padding": "global_block_aligned_row_stride",
      "train_shards": [
        1
      ],
      "train_tokens": 348019616,
      "val_shards": [
        2,
        3,
        4,
        5,
        6,
        7,
        8,
        9,
        10,
        6542
      ],
      "val_tokens": 3425453752,
      "vocab": 8192
    },
    "experiment": {
      "bucket_edges": [
        0,
        1,
        2,
        3,
        4,
        5,
        6,
        11,
        21,
        51,
        101,
        201,
        501,
        1001,
        5001
      ],
      "cpu_smoke": false,
      "device_batch_size": 72,
      "frequency_definition": "exact_train_epoch_context_count",
      "frequency_index_scope": "complete upstream train epoch before controlled block resampling",
      "frequency_key_type": "exact_context",
      "frequency_probe_batches": 2,
      "frequency_source_split": "train",
      "hash_bucket_occupancy_used_for_gap": false,
      "injection_norm_components": [
        "bigram",
        "trigram",
        "total"
      ],
      "injection_norm_definition": "sqrt(mean(hidden_state.float() ** 2))",
      "max_training_steps": 2000,
      "model_class": "GPT",
      "probe_detail_files": "probe_details/{step}_{split}.npz",
      "probe_detail_format": "npz_v1",
      "probe_detail_scope": "fixed_probe_batches",
      "probe_frequency_mode": "exact_context",
      "probe_steps": [
        100,
        200,
        300,
        400,
        500,
        600,
        700,
        800,
        900,
        1000,
        1200,
        1400,
        1600,
        1800,
        2000
      ],
      "seed": 42,
      "sequence_len": 2048,
      "torch_compile": false,
      "total_batch_tokens": 147456,
      "trace_all_batches": false,
      "trace_compression": true,
      "trace_manifest": "trace_manifest.jsonl",
      "trace_root": "/data4/guoshaoyang/ngram-gap-lab/data/runs_fixed/ngram5_order5_trigram_fixed/batch_trace",
      "trace_schema": "batch_trace_npz_v1",
      "trace_validation_mode": "fixed_validation_batches_every_step",
      "train_probe_batch_hashes": [
        "7a5f4f2df1f40dd8bf9a39ea8c4a42b557825fa46336ca1ce5b33d7fc1c23a17",
        "36ade48f85b944d3efa80dc290a0fd4615be4ddf78b5929e40c54cbc62e54876"
      ],
      "trunk": "transformer",
      "val_probe_batch_hashes": [
        "d8b5e09aeaf988ddc7d5418b53d283387154c9b0dc16d7f21d5a7e51d3662046",
        "fa4b4dcaf5ceaa462caa5823317a91603a525ba071a968246c2ed25e2c93f8f9"
      ],
      "validation_batch_hashes": [
        "d8b5e09aeaf988ddc7d5418b53d283387154c9b0dc16d7f21d5a7e51d3662046",
        "fa4b4dcaf5ceaa462caa5823317a91603a525ba071a968246c2ed25e2c93f8f9",
        "12406ce72a4635f24651be75771fd1d5b1b307926b69a345b8457a4a0c72a037",
        "3f8670597b92bd02825502bab9db085a174539be36ec0f9953d171a1d72841ba"
      ],
      "validation_batches": 4,
      "validation_interval_steps": 10
    },
    "model_provenance": {
      "note": "loaded from the repository model definition",
      "source": "/data4/guoshaoyang/ngram-gap-lab/code/train.py",
      "source_description": "repository NanoGPT with n-gram injection tables",
      "trunk": "transformer"
    },
    "optimizer": {
      "adam_betas": [
        0.8,
        0.95
      ],
      "class": "MixedOptimizer",
      "grouping": "nanogpt",
      "learning_rate": 0.004,
      "matrix_optimizer": "adamw",
      "ngram_optimizer": "mixed",
      "ngram_ve_betas": [
        0.0,
        0.99
      ],
      "ngram_ve_lr_scale": 2.0,
      "weight_decay": 0.1
    },
    "parameter_counts": {
      "total": 1278819840
    },
    "schedule": {
      "adam_warmdown_ratio": 0.65,
      "final_lr_fraction": 0.05,
      "ngram_lr_schedule": "constant",
      "warmup_ratio": 0.0
    },
    "schema_version": 1
  },
  "schema_version": 1,
  "status": "complete",
  "total_seconds": 534.3421177864075,
  "total_tokens": 294912000,
  "training_seconds": 453.317990064621
}