config.json

{
  "architecture": "BashkirRoBERTa Pre-LayerNorm encoder with a linear token-classification head",
  "task": "person_named_entity_recognition",
  "labels": ["O", "B-PER", "I-PER"],
  "vocab_size": 16384,
  "hidden_size": 640,
  "num_hidden_layers": 8,
  "num_attention_heads": 10,
  "intermediate_size": 2560,
  "max_position_embeddings": 256,
  "bos_token_id": 2,
  "eos_token_id": 3,
  "pad_token_id": 0,
  "tokenization": "SentencePiece per whitespace-delimited word; predict at the first piece",
  "output": "word-level BIO tags from logits [batch, sequence, 3]",
  "default_model": "int8",
  "models": {
    "int8": "onnx/model_int8.onnx",
    "fp32": "onnx/model_fp32.onnx"
  }
}