# ⚒️ ForgeX v3.2.0-open-source

中文 | English

ForgeX 是一个本地运行的 AI 模型训练、数据处理、蒸馏、模型编辑与部署工具箱。它把常见的大模型训练流程集中到一个 Gradio UI 里，适合个人研究、课程实验、模型微调、数据集清洗、LoRA 训练、GGUF/Ollama 导出与原型验证。

开源版默认不需要 license.key。原用户版中的轻量混淆模块已经还原为可读源码，便于审计、学习和二次开发。

这个项目是做什么的？

ForgeX 的目标是把“大模型训练工坊”做成本地可视化工具：你可以导入数据集，选择本地或 Hugging Face 模型，配置训练参数，运行 SFT/DPO/LoRA/QLoRA 训练，做 API 教师蒸馏，处理文档课程数据，尝试多模态训练，最后把模型合并、量化并导出到 GGUF 或 Ollama。

主要功能

• 数据集管理：上传、预览、删除、格式转换、清洗、去重、质量分析。
• 标准训练：SFT / DPO / LoRA / QLoRA，带智能参数推荐与显存估算。
• API 教师蒸馏：用 OpenAI 兼容接口或本地 Ollama 生成训练数据、批改答案、训练学生模型。
• 文档课程训练：解析 TXT / Markdown / PDF / DOCX / JSONL，生成课程式训练样本。
• 自我进化训练：从主题生成指令、生成候选回答、筛选偏好数据并迭代训练。
• 多模态训练：面向图文/音文理解的 LLaVA-style 原型训练流程。
• 模型手术：深度扩展、宽度扩展、词表扩展、知识嫁接、MoE upcycle。
• 模型合并与导出：LoRA 合并、GGUF 转换、Ollama Modelfile 生成。
• 模型编辑器：模型身份、系统提示词、知识库、安全护栏、推理参数与模型卡维护。
• 环境适配：启动器自动创建 .venv，并根据 NVIDIA/CPU 环境安装 PyTorch。

快速开始

1. 准备环境

需要 Python 3.10+。建议放在较短且可写的路径中，例如 Windows 下的 C:\ForgeX。

2. 启动

Windows：

run.bat

Linux / macOS：

chmod +x run.sh install.sh install_forgex.sh
./run.sh

首次运行会自动创建 .venv 并安装依赖。UI 默认监听：

http://127.0.0.1:7860

3. 常用命令

python launcher.py run             # 安装依赖并启动 UI
python launcher.py install         # 只安装/更新运行环境
python launcher.py fingerprint     # 输出本机指纹（兼容旧授权机制）
python launcher.py license-info    # 显示旧 license 信息；开源版无需 license
python selfcheck.py                # 静态自检：Python 文件可编译、核心接口存在
python smoke_test.py               # 轻量冒烟测试

项目结构

ForgeX/
├─ core/                    # 训练、蒸馏、导出、数据、模型编辑等核心模块
├─ core/simulation/         # 训练/显存估算与模拟工具
├─ data/configs/            # 默认配置
├─ data/datasets/           # 用户数据集目录（默认不入库）
├─ data/loras/              # 训练输出目录（默认不入库）
├─ data/models_cache/       # 模型缓存目录（默认不入库）
├─ docs/                    # 架构、路线图、开源说明
├─ examples/                # 示例数据
├─ main.py                  # Gradio UI 主入口
├─ launcher.py              # 跨平台启动器与环境安装器
├─ requirements.txt         # Python 依赖
└─ constraints.txt          # UI 兼容性约束

注意事项

• 本仓库不包含模型权重，也不包含原用户包中的大型训练数据文件。
• 训练大模型可能需要大量显存、磁盘空间和下载流量。
• 使用第三方模型、数据集、API 或论文资料时，请自行遵守其许可证、服务条款和版权要求。
• 部分模型加载流程会使用 trust_remote_code=True。只加载你信任的模型仓库。
• API Key 只应通过 UI 或环境变量临时提供，不要提交到 Git。

开源发布前检查清单

• 核心混淆模块已还原为可读源码。
• 删除 license.key。
• 删除大型训练数据 train_*.jsonl。
• 删除日志、临时文件和本地模型缓存。
• 加入 .gitignore、LICENSE、贡献说明、安全说明、架构文档和示例数据。
• python -m py_compile 静态编译通过。

许可证

本开源准备包默认使用 MIT License。如你希望改成 Apache-2.0、GPL-3.0 或保持闭源商业授权，请在发布前替换 LICENSE。

---

⚒️ ForgeX v3.2.0-open-source

ForgeX is a local AI model training, dataset processing, distillation, model editing, and deployment toolbox. It brings common LLM workflows into a Gradio-based UI for research, learning, fine-tuning, dataset cleaning, LoRA training, GGUF/Ollama export, and rapid prototyping.

The open-source build does not require license.key. The lightweight packaging obfuscation from the user build has been converted back into readable source code for review, learning, and contribution.

What is ForgeX?

ForgeX is designed as a local “LLM forge.” You can import datasets, choose local or Hugging Face models, configure training parameters, run SFT/DPO/LoRA/QLoRA jobs, perform API teacher distillation, build document-based curricula, experiment with multimodal training, and finally merge, quantize, or export models to GGUF/Ollama.

Features

• Dataset management: upload, preview, delete, convert, clean, deduplicate, and analyze datasets.
• Standard training: SFT / DPO / LoRA / QLoRA with smart parameter recommendations and VRAM estimates.
• API teacher distillation: use OpenAI-compatible APIs or local Ollama to generate data, grade answers, and train student models.
• Document curriculum training: parse TXT / Markdown / PDF / DOCX / JSONL and generate curriculum-style training samples.
• Self-evolution training: generate instructions from topics, create candidates, build preference pairs, and iterate.
• Multimodal training: prototype LLaVA-style image-text/audio-text training workflows.
• Model surgery: depth expansion, width expansion, vocabulary expansion, knowledge grafting, and MoE upcycling.
• Merge and export: merge LoRA adapters, convert to GGUF, and generate Ollama Modelfiles.
• Model editor: manage identity, system prompts, knowledge base, safety guardrails, inference parameters, and model cards.
• Runtime adaptation: the launcher creates .venv and installs a PyTorch profile based on NVIDIA/CPU hardware.

Quick Start

1. Requirements

Python 3.10+ is required. A short writable path is recommended, for example C:\ForgeX on Windows.

2. Launch

Windows:

run.bat

Linux / macOS:

chmod +x run.sh install.sh install_forgex.sh
./run.sh

On first launch, ForgeX creates .venv and installs dependencies automatically. The default UI address is:

http://127.0.0.1:7860

3. Useful Commands

python launcher.py run             # install dependencies and launch UI
python launcher.py install         # install/update runtime only
python launcher.py fingerprint     # print machine fingerprint for legacy licensing
python launcher.py license-info    # show legacy license info; not required for OSS build
python selfcheck.py                # static self-check
python smoke_test.py               # lightweight smoke test

Repository Layout

ForgeX/
├─ core/                    # core training, distillation, export, dataset, and editing modules
├─ core/simulation/         # training/VRAM estimation utilities
├─ data/configs/            # default configuration
├─ data/datasets/           # user datasets; ignored by Git
├─ data/loras/              # training outputs; ignored by Git
├─ data/models_cache/       # model cache; ignored by Git
├─ docs/                    # architecture, roadmap, and open-source notes
├─ examples/                # sample datasets
├─ main.py                  # Gradio UI entry point
├─ launcher.py              # cross-platform launcher and runtime installer
├─ requirements.txt         # Python dependencies
└─ constraints.txt          # UI compatibility constraints

Important Notes

• This repository does not include model weights or the large training dataset from the original user package.
• Training large models may require significant VRAM, disk space, and network bandwidth.
• When using third-party models, datasets, APIs, or papers, you are responsible for following their licenses and terms.
• Some model loading paths use trust_remote_code=True. Only load model repositories you trust.
• Do not commit API keys, private datasets, generated model weights, or commercial license files.

Pre-release Checklist