model: {architecture: ali, d_enc: 256, d_main: 512, n_enc: 2, n_dec: 2, n_main: 6,
  n_heads_enc: 4, n_heads_main: 8, seq_len: 1024, dropout: 0.0, target_ratio: 6.0,
  max_chunk_mult: 3.0, dechunk: ema, router_noise: 0.0, rope_base: 10000.0}
memory:
  ali:
    budget_B: 65536
    staging_capacity: 262144
    policy: random
    value_mode: learned
    candidate_types: [surface]
    write_rule: mean
    eta_min: 0.0
    memory_scope: persistent
    consolidate_every: 50
    max_span: 64
    gate: vector
    retrieval: true
    read_dropout: 0.1
    gate_hidden: 0
    type_dim: 16
    gate_bias_init: -2.0
    probe_seqs: 2
    probe_reps: 8
    horizon: 32
    predictor: mlp
    lambda_util: 0.1
    enabled: true
    inventory_kwargs: {}
    probe_mix: null
    util_batch_slots: 1024
  pkm: {n_keys: 256, topk: 8, heads: 1, d_key: null, value_lr_mult: 1.0}
objective: {name: ratio, schedule: null, alpha: 0.03, alpha_parity_cap: 0.06, global_anchor: 0.0,
  premium_clip: 4.0, premium_kind: flores_per_sent, tau1: 0.4, tau2: 0.7, freeze_scope: router,
  cpc_decay: 0.9, genreg_target_entropy: null, genreg_eta: 0.05}
data:
  corpus: corpus
  langs: [en, tr, fi]
  mixture: null
  doc_filters: null
  byte_classes: scratch/data/byte_classes.json
training:
  seed: 0
  max_steps: 6000
  batch_size: 32
  grad_accum: 1
  lr: 0.0003
  min_lr_frac: 0.1
  warmup: 300
  weight_decay: 0.1
  grad_clip: 1.0
  betas: [0.9, 0.95]
  router_lr_mult: 1.0
  log_every: 50
  amp: bf16
  compile: false
  device: cuda
checkpointing:
  run_name: ali_mem-random_s0
  runs_dir: runs
  ckpt_every: 200
  keep: false
  milestones: [1000, 3000]
  resume: auto
  keep_schedule: log2
  keep_extra_offsets: [2, 4, 6, 8]
  hub_repo: AdaptiveChunking/{run_name}
  hub_private: false
  keep_local: false
evaluation:
  eval_every: 200
  eval_batches: 8
  probe_sents: 128
  probe_bpb_batches: 4
  log_spaced_probe: true
  dense_early_every: 0
  dense_early_until: 0
  dump_reps: true
  dump_reps_thresholds: [0.3, 0.4, 0.5, 0.6, 0.7]