spb/forge Public MIT

1{2  "_comment": "batch_size is the MICRO-batch (8 x 8 x 1024 = 65536 tokens/step); the autograd tape holds all activations. precision is parsed but not yet honored - all kernels are f32 (see README \"Not yet done\").",3  "model": {4    "name": "gpt-25m",5    "n_layers": 8,6    "d_model": 512,7    "n_heads": 8,8    "n_kv_heads": 8,9    "d_ff": 1408,10    "vocab_size": 8192,11    "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": 0.0006,22    "min_lr_ratio": 0.1,23    "warmup_steps": 2000,24    "max_steps": 50000,25    "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": 1000,34    "eval_every": 500,35    "eval_batches": 20,36    "seed": 133737  }38}39