veronica-pretrain-24L.json

{
  "vocab_size": 50271,
  "n_layer": 24,
  "n_head": 12,
  "n_embd": 768,
  "max_position_embeddings": 2048,
  "mlp_mult": 4,
  "num_funcs": 3,
  "router_dim": 128,
  "use_channel_attention": false,
  "dropout": 0.0,
  "layer_norm_epsilon": 1e-05,
  "router_aux_weight": 0.005,
  "use_flash_attn": true,
  "router_tau": 1.6,
  "rope_theta": 10000.0,
  "gradient_checkpointing": true,
  "_comment": "Ultra-deep reasoning config: 24L/12H/768d, mlp_mult=4.5 (~390M params). Maximum reasoning depth for RAG + multi-step inference. ~14-16GB VRAM."
}