config.json

{
  "hidden_size": 640,
  "num_hidden_layers": 8,
  "num_attention_heads": 10,
  "intermediate_size": 2560,
  "vocab_size": 16384,
  "max_position_embeddings": 256,
  "architecture": "BashkirRoBERTa Pre-LayerNorm encoder with a linear word-classification head",
  "task": "word_level_language_identification",
  "labels": [
    "O",
    "BA",
    "RU",
    "TAT",
    "OTHER"
  ],
  "bos_token_id": 2,
  "eos_token_id": 3,
  "pad_token_id": 0,
  "unk_token_id": 1,
  "tokenizer_file": "spm_bashkir_bert_16k.model",
  "tokenization": "Unicode word/punctuation boundaries; SentencePiece per token; first-subword alignment",
  "word_window_overlap": 16,
  "long_text": "Overlapping complete-word windows; select the window closest to its center",
  "input": "input_ids: int64 [batch, sequence]",
  "output": "logits: float [batch, sequence, 5]",
  "default_model": "int8",
  "models": {
    "int8": "onnx/model_int8.onnx",
    "fp32": "onnx/model_fp32.onnx"
  }
}