{
  "_comment": "forge-tiny-25m-smoke — same skeleton as forge-tiny-25m but on the existing TinyStories 4k-vocab data, sized for a fast end-to-end validation run (~55 min at ~16k tok/s f32 on an M5 Max). Purpose: prove the full loop (dataloader -> Muon+WSD -> qk_norm -> checkpoint -> resume -> eval -> generate) before spending hours on the real mix. 16 x 8 x 1024 = 131,072 tokens/step x 400 steps = 52.4M tokens (~2.7 epochs of the local 19.1M-token set). Loss should fall well below 3.0; sampled stories must be coherent English. If this run misbehaves, nothing bigger gets launched.",
  "model": {
    "name": "forge-tiny-25m-smoke",
    "n_layers": 8,
    "d_model": 384,
    "n_heads": 6,
    "n_kv_heads": 3,
    "d_ff": 1024,
    "vocab_size": 4096,
    "context_length": 1024,
    "tied_embeddings": true,
    "use_rope": true,
    "rope_theta": 10000.0,
    "norm": "rmsnorm",
    "norm_eps": 1e-06,
    "activation": "swiglu",
    "qk_norm": true,
    "dropout": 0.0
  },
  "train": {
    "optimizer": "muon",
    "muon_lr": 0.02,
    "muon_momentum": 0.95,
    "lr": 0.0005,
    "min_lr_ratio": 0.1,
    "schedule": "wsd",
    "wsd_decay_frac": 0.15,
    "warmup_steps": 40,
    "max_steps": 400,
    "beta1": 0.9,
    "beta2": 0.95,
    "eps": 1e-08,
    "weight_decay": 0.1,
    "grad_clip": 1.0,
    "batch_size": 16,
    "grad_accum_steps": 8,
    "precision": "f32",
    "checkpoint_every": 100,
    "eval_every": 100,
    "eval_batches": 20,
    "seed": 1337
  }
}