{
  "_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\").",
  "model": {
    "name": "gpt-25m",
    "n_layers": 8,
    "d_model": 512,
    "n_heads": 8,
    "n_kv_heads": 8,
    "d_ff": 1408,
    "vocab_size": 8192,
    "context_length": 1024,
    "tied_embeddings": true,
    "use_rope": true,
    "rope_theta": 10000.0,
    "norm": "rmsnorm",
    "norm_eps": 1e-06,
    "activation": "swiglu",
    "dropout": 0.0
  },
  "train": {
    "lr": 0.0006,
    "min_lr_ratio": 0.1,
    "warmup_steps": 2000,
    "max_steps": 50000,
    "beta1": 0.9,
    "beta2": 0.95,
    "eps": 1e-08,
    "weight_decay": 0.1,
    "grad_clip": 1.0,
    "batch_size": 8,
    "grad_accum_steps": 8,
    "precision": "f32",
    "checkpoint_every": 1000,
    "eval_every": 500,
    "eval_batches": 20,
    "seed": 1337
  }
}