"""Hugging Face configuration for the character n-gram OCR-quality scorer.

Uploaded to the Hub next to modeling_char_ngram.py; consumers load the model
with trust_remote_code=True (see that file's docstring).

AI Disclosure:
    Models:         Claude Fable 5 (claude-fable-5)
    AI-Generated:   fully         # fully | mostly | partially | none
    Human-Reviewed: none          # fully | partially | minimally | none
"""

from transformers import PretrainedConfig


class CharNgramConfig(PretrainedConfig):
    model_type = "char-ngram"

    def __init__(
        self,
        vocab_size=256,
        order=6,
        backoff_alpha=0.4,
        total_tokens=0,
        ngram_sizes=None,
        bos_token_id=0,
        eos_token_id=0,
        pad_token_id=0,
        unk_token_id=1,
        **kwargs,
    ):
        self.vocab_size = vocab_size
        self.order = order
        self.backoff_alpha = backoff_alpha
        self.total_tokens = total_tokens
        # number of distinct k-grams per order 1..order (buffer shapes for from_pretrained)
        self.ngram_sizes = ngram_sizes or [0] * order
        self.unk_token_id = unk_token_id
        self.num_hidden_layers = 1  # dummy: generation utilities expect it; no cache is used
        self.use_cache = False
        super().__init__(
            bos_token_id=bos_token_id, eos_token_id=eos_token_id, pad_token_id=pad_token_id, **kwargs)