SAVRN
Search Contact SAVRN

Open-weight model · Text generation

macbert4csc-base-chinese

by Ming Xu (徐明) shibing624/macbert4csc-base-chinese

macbert4csc-base-chinese evaluate SIGHAN2015 test data： 由于训练使用的数据使用了SIGHAN2015的训练集（复现paper），在SIGHAN2015的测试集上达到SOTA水平。 模型结构，魔改于softmaskedbert： 本项目开源在中文文本纠错项目：pycorrector，可支持macbert4csc模型，通过如下命令调用： 当然，你也可使用transformers调用： SIGHAN+Wang271K中文纠错数据集，数据格式：…

Parameters102M
Context512
Weights1.3 GB
Licenseapache-2.0
AccessOpen weights
Monthly Downloads543.5k

Runs On

What it takes to serve macbert4csc-base-chinese (102M parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.

PrecisionWeightsMemory neededCheapest setupPer hourAlso fits
16-bit 0.2 GB 0.2 GB 1x MI300X (192 GB)
Vultr		 $1.85 1x H100 $1.99 · 1x MI325X $2.00
8-bit 0.1 GB 0.1 GB 1x MI300X (192 GB)
Vultr		 $1.85 1x H100 $1.99 · 1x MI325X $2.00
4-bit 0.1 GB 0.1 GB 1x MI300X (192 GB)
Vultr		 $1.85 1x H100 $1.99 · 1x MI325X $2.00

Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Sep 18, 2026.

SAVRN's Notes on macbert4csc-base-chinese

When Chinese text arrives with wrong characters, this is the model that corrects them before anything downstream reads it. Ming Xu built it on the MacBERT approach in arXiv:2004.13922, trained it on the shibing624/CSC data, and supports it through the pycorrector project. Hardware is a non-issue: 102M parameters, 0.2 GB of weights and 0.2 GB of memory at 16-bit. The cheapest card in our Index, one MI300X with 192 GB at $1.85 per hour, would use a tenth of one percent of its memory, so plan it as a sidecar on shared hardware.

Apache 2.0 covers commercial deployment, modification and redistribution with an express patent grant, as long as notices stay intact and significant changes are stated. Check two things: the 512-token window means long documents are corrected in pieces, and the publisher used the SIGHAN2015 training set, so scores on that test set are not an external measure.

SAVRN Research, 2026-09-18

Model Card

By Ming Xu (徐明), published under apache-2.0, revision 615e6e09ef9a.

MacBERT for Chinese Spelling Correction(macbert4csc) Model

中文拼写纠错模型

macbert4csc-base-chinese evaluate SIGHAN2015 test data：

Correct-Precision Correct-Recall Correct-F1
Chararcter-level 93.72 86.40 89.91
Sentence-level 82.64 73.66 77.89

由于训练使用的数据使用了SIGHAN2015的训练集（复现paper），在SIGHAN2015的测试集上达到SOTA水平。

模型结构，魔改于softmaskedbert：

Usage

本项目开源在中文文本纠错项目：pycorrector，可支持macbert4csc模型，通过如下命令调用：

from pycorrector.macbert.macbert_corrector import MacBertCorrector

m = MacBertCorrector("shibing624/macbert4csc-base-chinese")

i = m.correct('今天新情很好')
print(i)

当然，你也可使用transformers调用：

Read the full model card (477 words)

Configuration

Architecture
BertForMaskedLM
Context length (tokens)
512
Layers
12
Hidden size
768
Feed-forward size
3,072
Attention heads
12
Vocabulary size
21,128
Model type
bert

Identity and Version

Repository
shibing624/macbert4csc-base-chinese
Publisher
Ming Xu (徐明)
Task
Text generation
Modality
Text
Library
transformers
Parameters
102M parameters
Languages
zh
Revision
615e6e09ef9a69ec487bc7c641ec3a311e2c11b9
First published
2022-03-02
Last updated
2025-09-15

Files and Weights

17 files, 1.3 GB in total. The weights are 3 files totalling 1.3 GB in bin, onnx, safetensors.

Weights3 files · 1.3 GB
Configuration6 files · 1.8 KB
Tokenizer5 files · 658.9 KB
Documentation1 file · 6.3 KB
Other1 file · 138.9 KB
Repository1 file · 1.2 KB
Every file
FileTypeSizeSHA-256
model.safetensorsWeights409.2 MB 73df11fd36fb
onnx/model.onnxWeights474.3 MB a4d9d6807c0c
pytorch_model.binWeights409.2 MB 37a03fe0f103
added_tokens.jsonConfiguration2 B —
config.jsonConfiguration659 B —
onnx/config.jsonConfiguration845 B —
onnx/generation_config.jsonConfiguration90 B —
onnx/special_tokens_map.jsonConfiguration125 B —
special_tokens_map.jsonConfiguration112 B —
README.mdDocumentation6.3 KB —
arch1.pngOther138.9 KB —
.gitattributesRepository1.2 KB —
onnx/tokenizer.jsonTokenizer439.1 KB —
onnx/tokenizer_config.jsonTokenizer394 B —
onnx/vocab.txtTokenizer109.5 KB —
tokenizer_config.jsonTokenizer330 B —
vocab.txtTokenizer109.5 KB —

License and Download

License
apache-2.0
Access
Open weights, no gate
Download size
1.3 GB
Download from Ming Xu (徐明)

Released by Ming Xu (徐明) through its official repository on Hugging Face. Read the license.

Built From

Memory Requirements

PrecisionWeights in memory
As published1.3 GB
16-bit0.2 GB
8-bit0.1 GB
4-bit0.1 GB

Weights only, from the published parameter count; the key-value cache and runtime add to this.

Compare macbert4csc-base-chinese

Questions About macbert4csc-base-chinese

How much GPU memory does macbert4csc-base-chinese need?

About 0.2 GB at 16-bit and 0.1 GB at 4-bit: the weights (102M parameters) plus a working margin. A long context needs more.

What is the cheapest GPU to run macbert4csc-base-chinese on?

At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.

Can I use macbert4csc-base-chinese commercially?

Yes. macbert4csc-base-chinese is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.

What is macbert4csc-base-chinese's context length?

512 tokens, from the maximum position embeddings in its published configuration.

Similar Models

Model · Text generation

SharperSwarm

Convergent Intelligence

SAGI (Swarm AGI) is a novel causal language model that integrates swarm intelligence dynamics with transformer architecture. The model treats cognition as a dynamic, adaptive system where multiple internal "agents" collaborate through differentiable routing, trust mechanisms, and shared memory. V3.2 introduces a revolutionary Self-Assessment Layer, allowing the system to predict its own performance, identify skill gaps, and autonomously design its own learning curriculum. 1. Pre-Assessment: Predict success, identify risks, recommend strategy. 2. Execution: Generate with selected strategy. 3. Real-Time Monitoring: Catch and correct errors during generation. 4. Post-Assessment: Update skill…

Open weights apache-2.0 103M parameters 1,024 tokens transformers
View model

Model · Text generation

pythia-70m-deduped

EleutherAI

The Pythia Scaling Suite is a collection of models developed to facilitate interpretability research (see paper). It contains two sets of eight models of sizes 70M, 160M, 410M, 1B, 1.4B, 2.8B, 6.9B, and 12B. For each size, there are two models: one trained on the Pile, and one trained on the Pile after the dataset has been globally deduplicated. All 8 model sizes are trained on the exact same data, in the exact same order. We also provide 154 intermediate checkpoints per model, hosted on Hugging Face as branches. The Pythia model suite was designed to promote scientific research on large language models, especially interpretability research. Despite not centering downstream performance as a…

Open weights apache-2.0 96M parameters 2,048 tokens transformers
View model

Model · Text generation

distilgpt2

DistilBERT community

DistilGPT2 (short for Distilled-GPT2) is an English-language model pre-trained with the supervision of the smallest version of Generative Pre-trained Transformer 2 (GPT-2). Like GPT-2, DistilGPT2 can be used to generate text. Users of this model card should also consider information about the design, training, and limitations of GPT-2. CONTENT WARNING: Readers should be aware this section contains content that is disturbing, offensive, and can propagate historical and current stereotypes. As the developers of GPT-2 (OpenAI) note in their model card, “language models like GPT-2 reflect the biases inherent to the systems they were trained on.” Significant research has explored bias and…

Open weights apache-2.0 88M parameters transformers
View model

Model · Text generation

ivieai_star_v1.0

George O. Uwaifo

This model is a fine-tuned version of GeorgeUwaifo/iviegpt2new01cresults on an unknown dataset. It achieves the following results on the evaluation set: The following hyperparameters were used during training: - learningrate: 5e-05 - trainbatchsize: 4 - evalbatchsize: 8 - gradientaccumulationsteps: 4 - totaltrainbatchsize: 16 - lrschedulertype: linear - lrschedulerwarmupsteps: 387 - numepochs: 5 - mixedprecisiontraining: Native AMP - Transformers 5.16.1 - Pytorch 2.11.0+cu128 - Datasets 4.8.5 - Tokenizers 0.23.1

Open weights mit 124M parameters transformers
View model