1experiment:2 seed: 13 model: configs/prisme_103m.yaml4 model_overrides:5 readout_zero_init: false6 readout_init_std: 0.27 ctrl_input: h8 ctrl_use_entropy: true9 init:10 backbone_checkpoint: experiments/runs/pretrain103m_seed01_20261029232706/checkpoint11 freeze_backbone: true12 auto_readout_scale: true13 data:14 kind: tokens15 train_bin: data/processed/fineweb_edu_300m/train.bin16 val_bin: data/processed/fineweb_edu_300m/val.bin17 budgets:18 - 019 - 120 - 221 - 422 - 823 budget_probs:24 - 0.225 - 0.226 - 0.227 - 0.228 - 0.229 budget_per: position30 train:31 batch_size: 1632 seq_len: 51233 steps: 300034 lr: 0.00135 warmup: 10036 weight_decay: 0.137 grad_clip: 1.038 log_every: 2039 eval_every: 50040 eval_batches: 1041 ckpt_every: 100042 final_exact_match: false43 name: gr_pilot_103m44 mode: global_recurrence45 baseline_kwargs:46 ffn_w: 256047model:48 name: prisme_103m49 vocab_size: 3276850 d_model: 76851 n_layers: 1252 n_heads: 1253 n_kv_heads: 354 ffn_width: 204855 max_seq_len: 409656 rope_theta: 10000.057 norm_eps: 1.0e-0558 tie_embeddings: true59 init_std: 0.0260 use_workspace: true61 w: 25662 ffn_w: 76863 n_slots: 1664 n_editor_blocks: 265 n_editor_heads: 466 r_max: 867 eta_max: 0.2568 clip_c: 1.069 clip_eps: 1.0e-0670 readout_zero_init: false71 readout_init_std: 0.272 readout_scale: 1.073 controller_hidden: 074 ctrl_input: h75 ctrl_use_entropy: true76 lambda_cost: 0.077 kappa_risk: 0.078 dense_writes: false79resolved_train:80 batch_size: 1681 seq_len: 51282 steps: 300083 lr: 0.00184 warmup: 10085 weight_decay: 0.186 grad_clip: 1.087 log_every: 2088 eval_every: 50089 eval_batches: 1090 ckpt_every: 100091