中文 AI 生成文本检测器（本科毕业设计最终版）

A fine-tuned BERT model that classifies Chinese text as either human-written (0) or AI-generated (1). The main released model is a document-level binary classifier; mixed-text boundary detection is an experimental extension provided by a separate span model.

📌 模型概述 / Overview

中文：本模型是基于 bert-base-chinese 微调的中文 AI 生成文本二分类器，为本科毕业设计「基于 BERT 微调的中文 AI 生成文本检测系统」的最终生产模型（v11c boundary-fix 版本）。当前主链路输出 Human / AI 二分类；[SEP] 边界标记与 Token 级 span detector 是配套的实验性扩展，用于探索构造型人机混写样本中的片段级分析。

English: A binary classifier fine-tuned on bert-base-chinese for Chinese AI-generated text detection. This is the final production checkpoint (v11c boundary-fix) of an undergraduate thesis project. [SEP] boundary markers and the token-level span detector are experimental extensions for constructed human/AI mixed-text analysis, not the default production inference path.

📊 评估指标 / Evaluation

Dataset	Samples	Accuracy	Precision	Recall	F1
Validation set	7,452	98.75 %	98.30 %	99.37 %	98.83 %
`core_v1_test_clean`	545	97.98 %	97.87 %	98.77 %	98.32 %
Independent eval (910)	910	98.57 %	93.08 %	98.67 %	95.79 %
Three-set average	—	98.56 %	—	—	—

The metrics above evaluate the document-level binary classifier. The historical token-level boundary result belongs to the separate chinese-ai-detector-span experimental model and should not be mixed with the main classifier metrics.

Independent eval by source (selected)

Source	Samples	Accuracy
Toutiao News (all)	377	100.0 %
Wikipedia CN	119	99.16 %
formal_collected	200	96.5 %
real_ai_gemini-3-pro-preview	24	100.0 %
real_ai_deepseek-v3.2	8	100.0 %

🏗️ 架构 / Architecture

Base model: bert-base-chinese (12 layers, hidden 768, 12 heads, vocab 21,128)
Head: BertForSequenceClassification (2 labels: 0 = human, 1 = AI)
Max sequence length: 256 tokens (train), 512 (supported)
Framework: transformers 4.57.3, PyTorch 2.0+
Parameters: ~102M

Training configuration

Setting	Value
Base model	`bert-base-chinese` (via `bert_v7_improved` intermediate checkpoint)
Train samples	63,113
Validation samples	7,452
Epochs	5 (best at epoch 2)
Batch size	8 × 4 grad accum
Learning rate	1e-5
Label smoothing	0.05
Max length	256
Early stopping patience	2

Data changes vs. v10 baseline

Removed 750 hard patterns + 1,767 unapproved samples + 7 length violations
Added 300 formal-collected weak-domain samples
Added 300 Llama-405B weak-domain samples
Added 2,131 long-AI boundary-fix samples (the key v11c contribution)
Net change: +207 rows vs. v10

🚀 使用方法 / Usage

python

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

MODEL_ID = "AnxForever/chinese-ai-detector-bert"
TEMPERATURE = 0.8165  # Temperature scaling, calibrated on 910 samples (ECE=0.0034)

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForSequenceClassification.from_pretrained(MODEL_ID)
model.eval()

text = "这是一段需要检测的中文文本。"
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=256)

with torch.no_grad():
    logits = model(**inputs).logits
    # Apply temperature scaling for calibrated confidence
    probs = torch.softmax(logits / TEMPERATURE, dim=-1)[0]

pred_idx = int(probs.argmax())
label = model.config.id2label[pred_idx]   # "human-written" or "AI-generated"
print(f"{label}  (confidence: {probs[pred_idx].item():.2%})")

Label mapping

0 → human-written (人类撰写)
1 → AI-generated (AI 生成)

Note on Temperature Scaling: T = 0.8165 was calibrated on a held-out 910-sample set and brings ECE from 0.0121 down to 0.0034. For uncalibrated probabilities, set TEMPERATURE = 1.0.

🎯 技术贡献 / Contributions

Data-centric risk governance The v11c model keeps the BERT backbone fixed and improves robustness through data cleaning, weak-domain supplementation, long-AI supplementation, and calibrated inference.
[SEP] boundary-marker experiment In constructed C2-style mixed samples, [SEP] was used as an explicit boundary hint between known human and AI segments. This is an engineering experiment for mixed-text modeling, not a claim that [SEP] itself can identify authorship without labels.
Two-stage experimental extension
- Stage 1: this model — document-level Human / AI classification
- Stage 2: separate span detector — token-level Human / AI tagging on mixed-text samples
- See AnxForever/chinese-ai-detector-span
Long-AI boundary-fix (v11c) 针对长 AI 段落在边界处易被误判的问题，补充 2,131 条长 AI 边界样本，使 256+ token 桶的准确率恢复到 V10 水平。

Note on mixed-text boundary detection

The boundary module was trained on a relatively small constructed mixed-text set. It is useful for demonstration, teaching, and secondary development, but it should be treated as an experimental prototype. For real business scenarios, mixed human/AI data from the target domain should be collected, labeled, retrained, and evaluated before deployment.

⚠️ 局限性 / Limitations

仅针对中文文本；对英文或其他语言无保证。
训练语料偏新闻/百科/技术/正式文体，对诗歌、古文、社交媒体短文本可能欠拟合。
当前默认发布能力是篇章级二分类；人机混写边界定位属于实验性扩展，不建议直接作为商业审核结论。
训练数据主要来自 DeepSeek、Gemini、GPT、Llama-405B 等主流模型；对经过重度改写的 AI 文本仍有遗漏风险。
对短文本、强人工改写文本、多次交替混写文本和目标域外文本，不保证固定准确率。

🗂️ 相关资源 / Related

📊 训练数据集 / Dataset: AnxForever/chinese-ai-detection-dataset
🎯 边界检测器 / Span detector: AnxForever/chinese-ai-detector-span

📜 License

MIT License

✍️ Citation

bibtex

@misc{anxforever2026chineseaidetectorbert,
  title  = {Chinese AI-Generated Text Detector with Boundary Markers (BERT v11c)},
  author = {AnxForever},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/AnxForever/chinese-ai-detector-bert}},
  note   = {Undergraduate thesis project}
}