benchmark/float/resolved-config.json

{
  "seed": 17,
  "mode": "float",
  "device": "cuda",
  "model": {
    "vocab_size": 4096,
    "context_length": 256,
    "d_model": 256,
    "n_layers": 6,
    "n_heads": 8,
    "ff_multiplier": 4,
    "dropout": 0.1,
    "activation_threshold": 0.5,
    "weight_threshold": 0.5
  },
  "data": {
    "train_bin": "data/full/train.bin",
    "validation_bin": "data/full/validation.bin",
    "tokenizer": "data/full/tokenizer.json"
  },
  "train": {
    "batch_size": 64,
    "max_steps": 100,
    "learning_rate": 0.0003,
    "min_learning_rate": 3e-05,
    "warmup_steps": 20,
    "weight_decay": 0.1,
    "grad_clip": 1.0,
    "eval_interval": 100,
    "eval_batches": 10,
    "checkpoint_interval": 100,
    "output_dir": "artifacts/benchmark",
    "log_interval": 10,
    "precision": "bf16"
  }
}