spb/forge Public MIT

Forge — LLM training from scratch in pure C++20 + Metal on Apple Silicon.

C++ 61.2% C 23% Python 7.6% TeX 7.2% CMake 1.1%

916 B · 39 lines json

Raw Blame History

1{2  "_comment": "Model 'richard' (10M params) — one epoch over the 19.14M-token TinyStories set: 584 steps x 64 sequences x 512 tokens = 19.1M tokens. Warmup is 10% of the run.",3  "model": {4    "name": "richard",5    "n_layers": 6,6    "d_model": 384,7    "n_heads": 6,8    "n_kv_heads": 6,9    "d_ff": 1024,10    "vocab_size": 4096,11    "context_length": 512,12    "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": 58,24    "max_steps": 584,25    "beta1": 0.9,26    "beta2": 0.95,27    "eps": 1e-08,28    "weight_decay": 0.1,29    "grad_clip": 1.0,30    "batch_size": 64,31    "grad_accum_steps": 1,32    "precision": "f32",33    "checkpoint_every": 200,34    "eval_every": 100,35    "eval_batches": 20,36    "seed": 133737  }38}39