Repository Public MIT
1{2 "_comment": "MobileLLM-style deep-and-thin variant of gpt-50m: 20 layers x d_model 448 (aspect ratio ~22 vs 64 for the wide config), head_dim 64. ~50.6M params. Same data/steps/optimizer as gpt-50m so the only variable is the shape.",3 "model": {4 "name": "gpt-50m-deep",5 "n_layers": 20,6 "d_model": 448,7 "n_heads": 7,8 "n_kv_heads": 7,9 "d_ff": 1216,10 "vocab_size": 4096,11 "context_length": 1024,12 "tied_embeddings": true,13 "use_rope": true,14 "rope_theta": 10000.0,15 "norm": "rmsnorm",16 "norm_eps": 1e-06,17 "activation": "swiglu",18 "dropout": 0.019 },20 "train": {21 "lr": 0.0005,22 "min_lr_ratio": 0.1,23 "warmup_steps": 117,24 "max_steps": 1170,25 "beta1": 0.9,26 "beta2": 0.95,27 "eps": 1e-08,28 "weight_decay": 0.1,29 "grad_clip": 1.0,30 "batch_size": 8,31 "grad_accum_steps": 8,32 "precision": "f32",33 "checkpoint_every": 200,34 "eval_every": 100,35 "eval_batches": 20,36 "seed": 133737 }38}39