sampling_recipe.json

{
  "protocol": "recursive_block_multiply",
  "seed_operands": "seed + 91",
  "seed_offsets": "seed + 17",
  "base_cutoff": 2,
  "max_position_id": 512,
  "digit_order": "least-significant first",
  "runs": {
    "scratch_pretrain": {
      "updates": 15000,
      "batch_size": 64,
      "seed": 42,
      "mix_weights": [
        0.4,
        0.3,
        0.3
      ],
      "split_train_max": 8,
      "split_copy_max": 32,
      "combine_k_max": 16,
      "combine_r_max": 12,
      "learned_router": false,
      "exact_replay": false,
      "checkpoint": "checkpoints/recursive_block.pt (later overwritten by allk)",
      "notes": "First 15k steps from random init. Replay uses the current sampler; early combine / split encoding lived in salt_cert_recursive.py and may not be bit-identical."
    },
    "combine_finetune": {
      "updates": 10000,
      "batch_size": 64,
      "seed": 42,
      "mix_weights": [
        0.1,
        0.1,
        0.8
      ],
      "split_train_max": 8,
      "split_copy_max": 32,
      "combine_k_max": 16,
      "combine_r_max": 12,
      "learned_router": false,
      "exact_replay": false,
      "checkpoint": "runs/mult_transformer/recursive_block_combine.pt",
      "notes": "Combine-heavy finetune. Replay uses the current sampler."
    },
    "combine_k2_finetune": {
      "updates": 10000,
      "batch_size": 64,
      "seed": 42,
      "mix_weights": [
        0.1,
        0.1,
        0.8
      ],
      "split_train_max": 8,
      "split_copy_max": 32,
      "combine_k_max": 16,
      "combine_r_max": 32,
      "learned_router": false,
      "exact_replay": false,
      "checkpoint": "runs/mult_transformer/recursive_block_combine_k2.pt",
      "notes": "Same mix as combine_finetune with longer combine operands (r_max=32)."
    },
    "allk_finetune": {
      "updates": 10000,
      "batch_size": 64,
      "seed": 42,
      "mix_weights": [
        0.15,
        0.1,
        0.75
      ],
      "split_train_max": 8,
      "split_copy_max": 32,
      "combine_k_max": 16,
      "combine_r_max": 32,
      "learned_router": false,
      "exact_replay": false,
      "checkpoint": "checkpoints/recursive_block.pt",
      "notes": "Last length-gated checkpoint (SPLIT already in the split prompt). Init for the router run. Replay uses the current sampler."
    },
    "router_finetune": {
      "updates": 10000,
      "batch_size": 64,
      "seed": 42,
      "mix_weights": [
        0.25,
        0.4,
        0.35
      ],
      "split_train_max": 12,
      "split_copy_max": 32,
      "combine_k_max": 16,
      "combine_r_max": 32,
      "learned_router": true,
      "exact_replay": true,
      "checkpoint": "checkpoints/recursive_block_router.pt",
      "notes": "10k finetune from recursive_block.pt. Same code path as train_recursive after the learned-router change. This is the dump that matches the public 8x8 97.7% checkpoint."
    }
  },
  "regenerate": "python export_training_data.py --output data/hf"
}