1name: pretrain103m2seed: 13model: configs/prisme_103m.yaml4data:5 train_bin: data/processed/fineweb_edu_300m/train.bin6 val_bin: data/processed/fineweb_edu_300m/val.bin7train:8 batch_size: 329 seq_len: 51210 steps: 6000 # 32*512*6000 = 98.3M tokens (actual pilot budget, see DECISIONS D13)11 lr: 6.0e-412 warmup: 30013 weight_decay: 0.114 grad_clip: 1.015 log_every: 2016 eval_every: 50017 eval_batches: 2018 ckpt_every: 100019