config.json

{
  "model_type": "llama",
  "vocab_size": 32000,
  "d_model": 768,
  "n_layers": 16,
  "n_heads": 12,
  "d_ff": 2048,
  "max_seq_len": 2048,
  "rope_theta": 10000.0,
  "rope_yarn": true,
  "rope_yarn_scale": 4.0,
  "rope_yarn_alpha": 1.0,
  "rope_yarn_beta": 32.0,
  "rms_norm_eps": 1e-5,
  "tie_word_embeddings": true,
  "torch_dtype": "bfloat16",
  "repetition_penalty": 1.2,
  "bos_token_id": 2,
  "eos_token_id": 3,
  "pad_token_id": 1
}