config.json

{
  "bos_token": "<s>",
  "eos_token": "</s>",
  "layer_norm_epsilon": 1e-05,
  "multi_query_attention": true,
  "quantization_bits": 4,
  "quantization_group_size": 128,
  "quantization_type": 1,
  "unk_token": "<unk>"
}