ronghua/xiaoyu-chinese-ai-detect

AI model by ronghua

From the publisher

Model card & documentation

Source preview

Read the publisher’s intended use, setup instructions, evaluations and limitations. The original model card is the source of truth.

当前模型版本：V3.3 一个中文「AI 生成 / 人写」判别模型。输入一段文本，输出一个整篇AI概率，以及每句话的AI概率（仅供参考）。底座为中文 RoBERTa（BERT-base 结构，12 层 / 768 hidden / 词表 21128），单 logit 输出， sigmoid 后得到「这段文本像 AI 写的」概率。同一份权重打包成三档体积，从 390 MiB 到 60 MiB。int8 与 fp32 判别力基本一致；int4 概率排序仍然保持，但边界样本的概率会明显漂移——具体读数见 quant/benchreadout.json（那份在仓库内可复现）。三档判决，只此一套线：显示的数字和判定用的线必须是同一个标准，不要自己另设阈值，否则会出现「45% 却判疑似 AI 参与」这种看不懂的输出。模型硬上限 512 token（中文约 600 汉字），超过这个长度必须切窗再合并。本仓库提供两种合并方式：不满 500 字符的文本切不出第二个窗，自动退化为单窗——此时整段一次前向（超 512 token 会被截断），两种口径给出同一个数。到 500 字符以上才开始切第二个窗，所以「单窗 / 多窗」的分界是 500 字符，不是 512 token。参考实现（与下方精度读数同一套口径，可直接抄）： len 口径里的 len 是窗口字符数，不是 token 数；相邻窗重叠的 100 字在每个窗里各计一次，不做去重——所以要加权就直接按上式算，别自己先合并重叠段再算。内部混合测试集（人工撰写文本 + 多来源 AI 生成文本）上的读数。三列的白话意思。前两列的「线」都设在 50%，意思是「AI生成概率 ≥ 50% 就算作判为像 AI」，也就是至少落到上面「疑似AI参与」这一档（50%–90%），而不是只算「AI特征明显」那一档（≥90%）：这个数字越低越好。这个数字越高越好。换个问法就清楚了：不问「0.5 这条线怎么样」，而问「如果要保证 100 篇 AI 稿里至少 99 篇被认出来，这条线得放到多松？放宽之后，会连带冤枉多少篇真人稿？」之所以非得这么比：等权和加权算出来的概率刻度不一样，同一个 0.5 对两者不是同一件事。直接拿 0.5 去比「冤枉率」，比出来的是刻度差异，不是判别力差异。只有先把「认出 AI 的能力」拉齐到同一水平（都是 99%），再比「为此付出多少冤枉真人的代价」，才是在比算法本身。要认出 99% 的 AI 稿，等权需要把线压到约 0.52，加权需要把线压到约 0.68。这两个线同样来自上面那批内部标定数据，不是你机器上测出来的，换语料要重新标定。还有两点请留意：冤枉率会比 1.36% 更低。这里故意用宽口径，不往好看的方向挑线。…

ronghua/xiaoyu-chinese-ai-detect on Hugging Face A short preview, not the full document.

Read the full model card ↗ · Preview checked 2026-11-09T09:00:07.494Z

Inside the original model card — Document outline

Headings are captured from the source. Links open the publisher’s document, not a locally hosted copy.

Documentation belongs to its respective authors. A listing is not a grant of reuse or training rights. Confirm the document’s own terms at the source.

Model overview

ronghua/xiaoyu-chinese-ai-detect is a task-unspecified model repository published by ronghua on Hugging Face. A library was not reported.

This page summarizes Hub metadata. For intended use, training data, evaluation results and limitations, consult the original model card.

Model facts

Task: Not reported
Library: Not reported
Recent downloads (30 days): 0
Cumulative likes: 0
Hugging Face trending score: 0
Reported safetensors parameters: 102,268,417
Architecture: BertForSequenceClassification
License: Not reported — check the model card
Access: Not gated by Hugging Face
Created: 2026-11-09T06:35:31.000Z
Last modified: 2026-11-09T06:40:12.000Z

Compatibility and lineage

Language coverage was not reported.

Base-model lineage was not included.

Parameter count is not a RAM/VRAM requirement. Check precision, quantization, context length and runtime compatibility in the model card; no hardware or API-cost claim is inferred here.

Use and evaluate

Check license terms and access requirements first. Review the model card’s documented loading instructions, evaluations and safety limitations. Benchmark scores are not imported by this directory, and popularity is not an accuracy ranking.

No model weights are downloaded or executed by AltAPIs. Never enable remote model code without reviewing it.

Model card and usage instructions

Model files and configuration

Community discussion

Source and freshness

Source: Hugging Face Hub. Metadata observed 2026-11-09T06:55:49.891Z. Daily imports are snapshots, not real-time monitoring.

Popularity and source listings do not establish security, suitability, licensing rights or benchmark performance.

Open original source