{
  "model_type": "nanochat-lfm2-quantum-gpt",
  "architecture": "LFM2-style hybrid quantum GPT (nanochat.gpt.GPT)",
  "stage": "sft",
  "base_model": "MarkChenX/lfm2-quantum-128m",
  "checkpoint_step": 849,
  "val_bpb": 0.4337037826104738,
  "sequence_len": 1024,
  "vocab_size": 32768,
  "n_layer": 16,
  "n_head": 16,
  "n_kv_head": 8,
  "n_embd": 1024,
  "head_dim": 64,
  "mlp_type": "quantum",
  "quantum_num_qubits": 4,
  "quantum_depth": 2,
  "window_pattern": "L",
  "mixer_pattern": "CCCCCCCCCCAAAAAA",
  "conv_kernel": 3,
  "rope_theta": 1000000.0,
  "ffn_pattern": "",
  "use_value_embeddings": false,
  "sft_data_mixture": "SmolTalk (general chat) + MMLU auxiliary_train x3 epochs + GSM8K train x4 epochs",
  "note": "This is NOT a transformers AutoModel config -- the architecture is custom (see bundled nanochat/gpt.py). Load with nanochat.checkpoint_manager.build_model() as shown in inference.py."
}