spb/forge Public MIT

1{2  "_comment": "~52M params, sized between gpt-25m and gpt-100m. vocab 4096 to match the local tok4096 train.bin (19.14M tokens). batch_size is the MICRO-batch: 8 x 8 x 1024 = 65536 tokens/step; 1170 steps = ~4 epochs over the 19.14M-token set. Warmup is 10% of the run. precision is parsed but not yet honored - all kernels are f32.",3  "model": {4    "name": "gpt-50m",5    "n_layers": 10,6    "d_model": 640,7    "n_heads": 10,8    "n_kv_heads": 10,9    "d_ff": 1728,10    "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