1seed: 12model: configs/prisme_103m.yaml3model_overrides:4 readout_zero_init: false5 readout_init_std: 0.26 ctrl_input: h7 ctrl_use_entropy: true8init:9 backbone_checkpoint: experiments/runs/pretrain103m_seed01_20261029232706/checkpoint10 freeze_backbone: true11 auto_readout_scale: true12data:13 kind: tokens14 train_bin: data/processed/fineweb_edu_300m/train.bin15 val_bin: data/processed/fineweb_edu_300m/val.bin16budgets:17- 018- 119- 220- 421- 822budget_probs:23- 0.224- 0.225- 0.226- 0.227- 0.228budget_per: position29train:30 batch_size: 1631 seq_len: 51232 steps: 300033 lr: 0.00134 warmup: 10035 weight_decay: 0.136 grad_clip: 1.037 log_every: 2038 eval_every: 50039 eval_batches: 1040 ckpt_every: 100041final_exact_match: false42name: gr_pilot_103m43mode: global_recurrence44baseline_kwargs:45 ffn_w: 256046