公开仓库地址 / Repository:
https://github.com/kevin0818-lxd/deepseek-text-humanizer
面向长文本的 全本地 英文人性化改写工作台。一键启动即可体验:自动分块、多路径语义迁移、分层评分、源锚定保真修复,并支持结构化评分报告与整稿导出。推理链完全跑在本地硬件(Apple Silicon / MLX),不依赖任何外部商用 API。
Fully local-first long-form English humanization workspace. One command boots a complete end-to-end flow: auto chunking, multi-skew semantic rewriting, stratified scoring with source-anchored drift repair, and a workbench with structured reports + whole-draft export. Every rewrite runs locally on your Apple Silicon GPU (MLX) — no external commercial APIs required.
| 维度 | 说明 |
|---|---|
| 三模型分层协作 | Mamba 130M 极速快写 → Gemma 3-12B 高质量定稿 → TinyLlama 1.1B 批评校正,三种底座 adapters 独立分发 |
| 零外部 API 依赖 | 评分、翻译、分块、改写、批评 — 全部本地完成,开箱即用 Demo 模式无需任何密钥 |
| 10 维确定性评分 | Human Naturalness · AI Trace Risk · Semantic Fidelity · Lexical Diversity · Syntactic Complexity 等全量启发式指标 |
| Wavelet 分块 + Skew Label | 小波信号自动选择 left/middle/right skew,Detector-Aware Polish 与 Local Rescue 双候选路径择优 |
| 语义网 + 句法树 3D | 可视化 Rewrite 前后语义网络节点、关系边与单句句法层级变化 |
| 一键可运行 | bash scripts/start_local_demo.sh 拉起前端 3000 / 后端 8000 |
| 源锚定漂移修复 | 序数、年份、专业术语、引述短语在 Output Layer 强制回正,拒绝数字丢失 |
| Mock → MLX 无感切换 | Demo 模式下所有写路径走本地确定性 mock;下载 adapters 重启即自动切到真 MLX 推理 |
bash scripts/start_local_demo.sh首次启动默认进入安全 Demo Mock 模式,全部用户侧主功能可直接体验:
- 手机
18800000000+ 验证码123456登录(本地 dev 模式) - 创建工作区 → 粘贴长英文原文 → 触发 Humanize
- 自动轮询进度 → 每个 chunk 给出 3 skew 候选并 auto-select
- 生成评分面板(10 指标 × before/after + composite + 3D 结构)
- 一键导出整稿
启动后访问 / After startup:
- Frontend:
http://127.0.0.1:3000 - Backend:
http://127.0.0.1:8000 - API Docs:
http://127.0.0.1:8000/docs
代码仓库不携带模型权重;三个底座维度的最新 adapters 通过 GitHub Releases 单独分发。启动脚本会自动探测并提示下载:
No released Gemma adapter found locally.
Release assets (three-model adapters):
Mamba adapter: https://github.com/kevin0818-lxd/deepseek-text-humanizer/releases/download/v1.0.0-adapters/mamba-130m-stratified-v1-adapter.safetensors
TinyLlama adapter: https://github.com/kevin0818-lxd/deepseek-text-humanizer/releases/download/v1.0.0-adapters/tinyllama-1.1b-stratified-v1-adapter.safetensors
Gemma adapter: https://github.com/kevin0818-lxd/deepseek-text-humanizer/releases/download/v1.0.0-adapters/gemma-3-12b-stratified-v1-adapter.safetensors
- 交互式下载(推荐):启动脚本询问时选
Y,三份 adapters 会被拉到data/models/adapter-release-v1/。 - 环境变量一键拉取:
CTWC_AUTO_PULL_ADAPTER=1 bash scripts/start_local_demo.sh
- 手动下载放置:
mkdir -p data/models/adapter-release-v1 cd data/models/adapter-release-v1 curl -L -O https://github.com/kevin0818-lxd/deepseek-text-humanizer/releases/download/v1.0.0-adapters/mamba-130m-stratified-v1-adapter.safetensors curl -L -O https://github.com/kevin0818-lxd/deepseek-text-humanizer/releases/download/v1.0.0-adapters/tinyllama-1.1b-stratified-v1-adapter.safetensors curl -L -O https://github.com/kevin0818-lxd/deepseek-text-humanizer/releases/download/v1.0.0-adapters/gemma-3-12b-stratified-v1-adapter.safetensors
| Adapter 资产 | 底座 Base Model | 推理角色 |
|---|---|---|
mamba-130m-stratified-v1-adapter.safetensors |
mlx-community/mamba-130m-hf-f32 |
Writer 极速快写路径(source-rewrite prompt 轮廓,抗中文回声) |
tinyllama-1.1b-stratified-v1-adapter.safetensors |
mlx-community/tinyllama-1.1b-chat-v1.0-4bit |
Critic 本地批评校正(detector-aware critic-reviser 信号) |
gemma-3-12b-stratified-v1-adapter.safetensors |
mlx-community/gemma-3-12b-it-4bit |
Writer 高质量定稿(left/middle/right skew 分层 stratification 训练) |
⚠️ 首次启用 MLX 模式时,MLX / Hugging Face 会在后台按 base model repo 名拉取底座权重;此过程仅发生一次。如果你的机器上已经缓存了对应底座,启动会直接进入推理。
Source (long-form English)
│
▼
Wavelet Chunker ──► Skew Label (left/middle/right)
│
▼
Coverage Stage ──► Mamba (fast) or Gemma (neutral)
│ └► source-anchored overlap + fidelity gate
▼
Style Stage ──► Detector-Aware Polish (Gemma)
│ └► AI-ish phrases / structural parallelism 降险
▼
Humanize Stage ──► (optional, wavelet flat/peaky 信号触发)
│
▼
Critic Stage ──► TinyLlama 本地批评 → drift repair
│
▼
Multi-Candidate Selection
│ └► summarize_local_humanizer_signals + topic-drift penalty
▼
Output Layer (cleanup + source-anchored fact/number/term repair)
│
▼
Scoring Workbench + Semantic Network 3D + Export
为保护隐私与合规,公开仓库不会包含以下内容:
- 真实
.env/.env.bak*/apps/web/.env.local文件及其中的任何密钥 - 外部商用 API(DeepSeek / Kimi / ZeroGPT)客户端、浏览器自动化脚本与对应测试
- 训练语料、基准数据集、激活码导出 CSV、用户上传 PDF
- 本地数据库(
*.db/*.sqlite3)、调试目录.dbg/、临时release-assets/与.public-sync/ - 个人截图
IMG_*.*、debug-*.md、运行日志
If you fork or clone this repository, never commit any of the above back upstream.
本项目以 Humanize 后在独立第三方检测器达到 0% AI 率 为目标工程指标,但任何最终产出在对外提交前,强烈建议做一次人工快速校对,重点关注:
- 提示词污染:检查是否存在
Re-written version:、Source sentence:、Chinese semantic representation:等脚手架标签。- 数字 / 专有名词漂移:对照原文快速核对年份、百分比、样本量、专有术语、人名地名机构名。
- 跨 chunk 衔接:长文的段落过渡处是否有重复的话语标记(However / Therefore / Furthermore 连续出现)或重叠句式。
Even with full local determinism and source-anchored repair, please do a quick manual review before any external submission, especially for prompt contamination, numeric drift, and cross-chunk discourse artifacts.
# 后端 (FastAPI + MLX)
cd apps/api
pip install -e ".[dev]"
python -m uvicorn app.main:app --host 127.0.0.1 --port 8000
# 前端 (Next.js 15 + React)
cd apps/web
npm install
npm run dev -- --hostname 127.0.0.1 --port 3000测试(核心回归):
cd apps/api
pytest tests/test_gemma_output_layer.py tests/test_scoring.py tests/test_machine_translation.py -qMIT