model: architecture: hnet_moe_router d_enc: 32 d_main: 48 n_enc: 1 n_dec: 1 n_main: 1 n_heads_enc: 2 n_heads_main: 2 seq_len: 128 dropout: 0.0 target_ratio: 6.0 max_chunk_mult: 3.0 dechunk: ema router_noise: 0.0 rope_base: 10000.0 moe_router: {n_experts: 5, gate: script, expert_families: null, common: inherit, default_family: latin, expert_init: clone, target_ratios: null, unilid_path: null, unilid_window: 32, unilid_temperature: 1.0, unilid_lang_experts: null, gate_temperature: 1.0, balance_coef: 0.01, detach_gate_in_ratio: true} memory: ali: budget_B: 512 staging_capacity: 2048 policy: utility value_mode: learned candidate_types: [surface, split, merge] write_rule: mean eta_min: 0.0 memory_scope: persistent consolidate_every: 5 max_span: 64 gate: vector retrieval: true read_dropout: 0.1 gate_hidden: 0 type_dim: 16 gate_bias_init: -2.0 probe_seqs: 1 probe_reps: 3 horizon: 32 predictor: mlp lambda_util: 0.1 enabled: true inventory_kwargs: {min_obs: 1, min_age: 0} probe_mix: null util_batch_slots: 1024 pkm: {n_keys: 16, topk: 4, 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, flex_targets: premium, flex_schedule: inverse_linear, flex_start_ratio: null, flex_ramp_start: 0.0, flex_ramp_end: 0.8, flex_freeze_at: null, flex_bits_source: flores_ngram, flex_bits_order: 5, flex_bits_anchor: en} data: corpus: synthetic langs: [en, de, hi] mixture: null doc_filters: null byte_classes: null training: seed: 0 max_steps: 20 batch_size: 4 grad_accum: 1 lr: 0.0003 min_lr_frac: 0.1 warmup: 2 weight_decay: 0.1 grad_clip: 1.0 betas: [0.9, 0.95] router_lr_mult: 1.0 log_every: 5 amp: none compile: false device: cpu checkpointing: run_name: smoke_moe-router-script runs_dir: /private/tmp/claude-501/-Users-suchirsalhan-Documents-PHD-PHD-beetle-hnet/3f003e5e-086f-4ff2-a93f-1f86ed43e9bd/scratchpad/runs ckpt_every: 10 keep: false milestones: [] 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: 10 eval_batches: 1 probe_sents: 4 probe_bpb_batches: 1 log_spaced_probe: true dense_early_every: 0 dense_early_until: 0 dump_reps: false dump_reps_thresholds: [0.3, 0.4, 0.5, 0.6, 0.7]