README.md

---
license: apache-2.0
language:
- zh
tags:
- chinese
- cws
- pos
- ner
- multi-task
- bert
library_name: pytorch
---

# BERTc-315M-MT

BERTc-315M-MT is a Chinese multi-task tagging model fine-tuned from
`Ismantic/BERTc-315M`. It predicts:

- CWS: Chinese word segmentation
- POS: PD-1998 POS tags mapped to the LTP tag set
- NER: Nh / Ns / Ni entity spans in BIES format

## Metrics

PD-1998 dev joint score:

- Joint score: **1.4712**
- CWS micro F1: **0.9840**
- POS per-word acc: **0.9800**
- NER micro F1: **0.9660**

Training recipe:

- backbone: `Ismantic/BERTc-315M`
- epochs: 5
- batch size: 64
- learning rate: `bert_lr=2e-5`, `head_lr=5e-4`
- warmup ratio: 0.1
- FGM: enabled, `eps=1.0`
- loss weights: `alpha_pos=2.0`, `beta_ner=0.5`

## Files

- `model.safetensors`: MT state dict for the backbone and task heads.
- `config.json`: BERTc backbone architecture.
- `model.py`: Modern BERTc implementation.
- `mt_model.py`: MT wrapper, tokenizer adapter, and decode helpers.
- `piece.model`: tokenizer model; load with `piece_tokenizer` using `cn_dict="no"`.
- `mask_token_id.txt`: mask token id used by the backbone tokenizer.
- `mt_config.json`: task metadata and source checkpoint.

## Usage

```python
from mt_model import BERTcForMT, PieceCharTokenizer

tok = PieceCharTokenizer(".")
model = BERTcForMT.from_pretrained(".")
out = model.predict("中华人民共和国万岁", tok)
print(out["words"])
print(out["pos"])
print(out["ner"])
```