BERTc-165M-CSC

BERTc-165M-CSC is a Chinese spelling correction model fine-tuned from Ismantic/BERTc-165M. It uses a Modern BERTc encoder with two heads:

• correction head: tied to the input embedding matrix
• detection head: binary error detection

Metrics

SIGHAN-15 sentence-level evaluation using the pycorrector-style 707-sample protocol:

• Sentence F1: 0.8308
• Accuracy: 0.8416
• Precision: 0.9516
• Recall: 0.7373
• TP/FP/FN/TN: 275 / 14 / 98 / 320

Training recipe:

• backbone: Ismantic/BERTc-165M
• epochs: 5
• batch size: 64
• learning rate: 5e-5
• warmup ratio: 0.1
• detection loss weight: 0.3
• inference threshold: 0.7
• max length: 128

Files

• model.safetensors: CSC state dict. cor_head.weight is intentionally omitted and tied to bert.embed.weight by csc_model.py.
• config.json: BERTc backbone architecture.
• csc_config.json: task and metric metadata.
• model.py: Modern BERTc implementation.
• csc_model.py: CSC wrapper and batch correction helper.
• piece.model: tokenizer model; load with piece_tokenizer using cn_dict="no".

Usage

from csc_model import BERTcForCSC, PieceCharTokenizer

tok = PieceCharTokenizer(".")
model = BERTcForCSC.from_pretrained(".")
texts = ["少先队员因该为老人让坐。"]
print(model.correct(texts, tok, threshold=0.7))

This is not a generative model. It performs same-length character replacement for Chinese spelling correction.