configs/config.json

{
 "input_dim": 512,
 "hidden_dim": 2048,
 "output_dim": 2048,
 "memory_dim": 512,
 "memory_slots": 64,
 "num_layers": 12,
 "num_associations": 8,
 "use_quantization": false,
 "use_mla": false,
 "mla_latent_dim": 32,
 "use_causal_lm": true,
 "vocab_size": 128000,
 "max_seq_len": 512,
 "causal_window_size": 64,
 "sae_k": 64,
 "ntm_memory_slots": 16,
 "d_model": 512,
 "mla_n_heads": 8,
 "mla_max_cache_len": 4096,
 "lm_num_attn_layers": 2,
 "lm_pooling": "mean",
 "_comment": {
  "input_dim": "输入维度(d_model)",
  "hidden_dim": "隐藏层维度(FFN)",
   "output_dim": "三脑内部输出维度(=hidden_dim, 非vocab_size)",
   "memory_dim": "记忆维度(d_mem)",
  "memory_slots": "记忆槽数(MEM_SLOTS)",
  "num_layers": "网络层数(ECN层)",
  "num_associations": "DMN关联头数",
  "use_quantization": "启用量化",
  "use_mla": "启用MLA",
  "mla_latent_dim": "MLA潜在维度",
  "use_causal_lm": "因果LM — next-token预测",
  "vocab_size": "128K多语言多领域词表(中/英/代码/数字/标点)",
  "max_seq_len": "最大序列长度",
  "causal_window_size": "因果窗口大小",
  "sae_k": "SAE稀疏度",
  "ntm_memory_slots": "NTM记忆槽数",
  "d_model": "模型维度(512)",
  "mla_n_heads": "MLA头数",
   "mla_max_cache_len": "MLA最大缓存长度",
   "lm_num_attn_layers": "CausalLM注意力层数(2-4)",
   "lm_pooling": "池化方式(mean/last)"
 },
  "_fix_log": "2026-06-13: 128K BPE tokenizer (train_128k_tokenizer.py)"
}