WeClone

🎉 What's Changed

WeClone 1.0.3 lets you build your life portrait from start to finish in the web UI!

Build portrait page

• A new Build portrait page checks your setup and lists anything missing with how to fix it. The build then runs in the background with progress for each stage. You can cancel it, and closing the page does not stop it.
• Chat records in the generic CSV format can be uploaded on the page as folders or zip files. In encrypted mode they are stored encrypted.
• The model service is chosen on the same page: the Codex model and effort, or for the API a provider, a key and a model from the list the service offers.
• Chat files are processed in parallel up to batch_size, so builds finish sooner.
• build-profile --csv-folder <dir> reads any folder of generic CSV chats.

Settings

• Model service: choose the backend and fill in its settings. Test connection checks the Codex sign-in, or lists the service's models and fills in the answer format and context window the service supports. Saving keeps the comments and order of settings.jsonc.
• Portrait versions: every build is listed. You can switch between them at any time and delete the ones you no longer need.

API backend

• New agent_distill_args options: context_window, reasoning_effort, extra_body and structured_output.
• The input limit follows the model's context window, 95% of it and 128K by default, so --max-context-tokens is no longer needed.

Web UI

• Error messages say what to do next, with the original error under Error details.

⚠️ Upgrade Notes

• DeepSeek is no longer detected by its address. For DeepSeek or another service that does not support json_schema, set structured_output to json_object. Choosing DeepSeek on the Settings page or running Test connection fills it in.
• If your API model's context window is smaller than 128K, set it under Settings → Model service.
• After updating, sign in to the web UI once more. Sign-in length under Settings returns to its default.

🐛 Bug fix

• Dropdowns now follow the page's theme and are readable in dark mode.
• A SOCKS proxy set in ALL_PROXY now works for the API backend. Run uv pip install -e . again after updating.

Full Changelog: https://github.com/xming521/WeClone/compare/v1.0.1...v1.0.3

😊 更新内容

WeClone 1.0.3 可以在网页上从头到尾完成人生画像的构建了！

构建画像页

• 新增「构建画像」页，先检查配置，列出缺少的设置和改法，然后在后台构建并显示每个阶段的进度。构建可以取消，关掉网页也不会停止。
• 通用 CSV 格式的聊天记录可以在这个页面以文件夹或 zip 上传，加密模式下加密存储。
• 模型服务也在这个页面选择：Codex 填模型和思考强度，API 选服务商、填 Key，再从服务列出的模型里选一个。
• 多个聊天文件按 batch_size 并行处理，构建更快。
• build-profile --csv-folder <目录> 可以读取任意通用 CSV 格式的聊天目录。

设置

• 模型服务：选择模型后端并填写设置。测试连接会检查 Codex 是否已登录，或列出服务的模型，并填好服务支持的回答格式和上下文长度。保存时保留 settings.jsonc 里的注释和顺序。
• 画像版本：列出每一次构建，可以随时切换，不要的版本可以删除。

API 后端

• agent_distill_args 新增 context_window、reasoning_effort、extra_body 和 structured_output。
• 输入上限按模型上下文长度的 95% 计算，默认 128K，不再需要 --max-context-tokens。

网页

• 报错信息会说明下一步怎么做，原始报错放在「错误详情」里。

升级说明

• 不再按接口地址识别 DeepSeek。DeepSeek 等不支持 json_schema 的服务需要把 structured_output 设为 json_object，在设置页选择 DeepSeek 或点测试连接会自动填好。
• API 模型的上下文长度小于 128K 时，请在「设置 → 模型服务」里填写。
• 升级后需要重新登录一次网页，登录有效期设置恢复为默认值。

问题修复

• 下拉菜单跟随网页主题，深色模式下也能正常阅读。
• API 后端支持 ALL_PROXY 里设置的 SOCKS 代理，升级后请重新运行 uv pip install -e .。

🎉 What's Changed

WeClone 1.0 is here! WeClone grows from fine-tuned clones into building a complete personal model of you. You can now distill a life portrait from your chat history, review it, and serve it over MCP as long-term memory for AI assistants or AI clones. The original fine-tuning workflow remains fully supported.

Life portrait

• weclone-cli build-profile builds your life portrait in one command, using Codex or any OpenAI-compatible API.
• The portrait is organized by domain and topic. Every fact has a privacy level and a time scope, and traces back to the chats it came from.
• Speaking style analysis: 179 style features in ten categories.
• The portrait is currently extracted in Chinese, and English chats have not been tested yet.

Assistant access

• Claude Code, Codex and other assistants read your portrait and domain summaries over MCP. You decide which domains and privacy levels they can read.

Web UI

• Review your portrait fact by fact, confirm domain summaries, view your speaking style, and manage access for assistants and chat bots.
• A new design with light and dark themes, available in Chinese and English.
• weclone-cli server --demo shows a demo portrait before you build your own.

Privacy

• Data is encrypted locally by default, with one password for the web UI and the command line.

Fine-tuned clone

• Chat bots such as AstrBot and LangBot connect to your clone with a token.
• Quantization settings now take effect in training and inference by @xhrxgr in #218
• GPUs without bfloat16, such as T4 and V100, are supported through the dtype option by @octo-patch in #212
• test-model asks you to start the server first when it is not running by @octo-patch in #213
• Generated training data is saved to the dataset set in the config by @Iams4kura in #229

⚠️ Upgrade Notes

• Update your settings.jsonc against the new settings.template.jsonc.
• New configs store data encrypted by default. Existing configs without security_args stay in plaintext.
• The life portrait, web UI and assistant access need only uv pip install -e .. Fine-tuning still needs --group main.

🐛 Bug fix

fix #85 fix #115

New Contributors

• @octo-patch made their first contribution in https://github.com/xming521/WeClone/pull/213
• @xhrxgr made their first contribution in https://github.com/xming521/WeClone/pull/218
• @Dessalines39394 made their first contribution in https://github.com/xming521/WeClone/pull/224
• @Londopy made their first contribution in https://github.com/xming521/WeClone/pull/232
• @Iams4kura made their first contribution in https://github.com/xming521/WeClone/pull/229
• @XiaoZheBrother made their first contribution in https://github.com/xming521/WeClone/pull/249
• @lhongjie70-sudo made their first contribution in https://github.com/xming521/WeClone/pull/253

Full Changelog: https://github.com/xming521/WeClone/compare/v0.3.03...v1.0.1

😊 更新内容

WeClone 1.0 正式发布！WeClone 现在从微调分身走向完整的个人模型建模。现在可以从聊天记录中总结出人生画像，经过审核后通过 MCP 作为长期记忆提供给 AI 助理或 AI 分身，原有的微调流程也会完整保留。

人生画像

• weclone-cli build-profile 一条命令生成人生画像，支持 Codex 或任意 OpenAI 兼容 API。
• 画像按板块与主题组织，每条事实标注隐私级别和时间范围，并可追溯原文。
• 语言风格分析：十大类、179 项风格特征。

助手接入

• Claude Code、Codex 等助手可通过 MCP 读取你的画像与板块总结，能读哪些板块、读到哪个隐私级别由你设定。

网页

• 逐条审核画像，确认板块总结，查看语言风格，管理助手和聊天机器人的接入。
• 全新设计，支持浅色和深色主题，提供中英文界面。
• weclone-cli server --demo 可以在生成自己的画像前先查看示例画像。

隐私

• 数据默认在本地加密存储，网页和命令行共用一个密码。

微调分身

• AstrBot、LangBot 等聊天机器人通过令牌接入分身。
• 量化配置在训练和推理中正确生效（@xhrxgr，#218）。
• 通过 dtype 选项支持 T4、V100 等不支持 bfloat16 的显卡（@octo-patch，#212）。
• test-model 在服务未启动时提示先启动服务（@octo-patch，#213）。
• 生成的训练数据保存到配置指定的数据集（@Iams4kura，#229）。

升级说明

• 请对照新的 settings.template.jsonc 更新 settings.jsonc。
• 新配置默认加密存储，未设置 security_args 的旧配置仍为明文。
• 人生画像、网页和助手接入只需 uv pip install -e .，微调仍需 --group main。

🎉 What's Changed

The key highlights of this update include an upgrade to Python 3.12 and optimization of the dataset pipeline.

Dependency and Environment Updates:

• Upgraded the required Python version to 3.12 in pyproject.toml and development settings, and updated the target version for linting and type checking to Python 3.12. [1] [2] [3]
• Updated dependencies: switched from a git-based install of llamafactory to a fixed version, added torchdata and torchaudio with CUDA 12.6 support, and refined platform-specific dependency markers for PyTorch packages. [1] [2]

Data

• Added the "<begin_chat>" marker in user messages, allowing for improved context in conversation flows.
• Updated the qa_generator.py to include a new mechanism for managing chat member relationships, allowing the addition of contextual information about the relationship between users in conversations.
• Refactored the CSV loading function to support loading user relationship data from a users.json file, improving the context provided during QA generation.
• Added a new configuration option add_relation to the dataset settings, enabling users to toggle this feature.

others

• Introduces OnlineLLM with thread‑pooled batch chat and optional JSON‑guided decoding; unifies JSON parsing across vLLM and OpenAI results.

• fix: fix triton source from default cuda129 to 126 by @MapleWithered in https://github.com/xming521/WeClone/pull/198

New Contributors

• @MapleWithered made their first contribution in https://github.com/xming521/WeClone/pull/198

Full Changelog: https://github.com/xming521/WeClone/compare/v0.3.02...v0.3.03

😊 更新内容

本次更新核心亮点包括升级至Python 3.12以及数据集管线优化。

依赖与环境更新：

• 在pyproject.toml和开发配置中将Python版本升级至3.12。
• 依赖项更新：将llamafactory从基于git的安装方式改为固定版本，新增支持CUDA 12.6的torchdata和torchaudio，并优化了PyTorch包的平台特定依赖标记。

数据处理

• 新增"<begin_chat>"标记，以提升对话流程的上下文连贯性
• 更新qa_generator.py，新增聊天成员关系管理机制，支持在对话中添加用户间关系的上下文信息
• 重构CSV加载函数，支持从users.json文件加载用户关系数据，增强问答生成时的上下文信息
• 在数据集配置中新增add_relation选项，允许用户自主启用/禁用此功能

其他

• 引入支持线程池批量聊天和可选JSON引导解码的OnlineLLM；统一了vLLM与OpenAI结果的JSON解析流程。

🎉 What's Changed

Enable configurable thinking in offline cleaning, improve image and gif handling in QA processing, refactor configuration models for cleaner dataset naming, and bump versions and dependencies for release v0.3.02

New Features:

• Introduce enable_thinking flag in LLMCleanConfig to control offline cleaning behavior
• Supporting scoring and cleaning of datasets containing images (assigning the highest score to QA pairs that include images).

Enhancements:

• Refactor cleaned_dataset_name to derive dynamically from original dataset
• Pass enable_thinking through vLLM inference pipeline and adjust repetition_penalty and max_new_tokens accordingly
• Implement CommonMethods to parse dataset names with modality-based suffixes and remove deprecated config fields

Build:

• Bump project version to 0.3.02 and config_version to 0.3.02
• Update dependencies: openai to 1.87.0, vllm to 0.10.0, torch to 2.7.1, add torchvision, transformers to 4.53.2, and triton to 3.3.1

Full Changelog: https://github.com/xming521/WeClone/compare/v0.3.01...v0.3.02

😊 更新内容

在离线清理中启用可配置的“思考”功能，改进问答处理中的图像和 GIF 处理，重构配置模型以实现更清晰的数据集命名，并为发布 v0.3.02 提升版本和依赖项。

新功能：

• 引入 enable_thinking 以控制离线清理行为
• 支持对含有图片的数据集打分清洗（含有图片的qa对赋值最高分）

改进：

• 重构 cleaned_dataset_name 以从原始数据集动态派生
• 将 enable_thinking 传递给 vLLM 推理管道，并相应调整 repetition_penalty 和 max_new_tokens
• 实现 CommonMethods 以解析带有模态后缀的数据集名称，并移除已弃用的配置字段

构建：

• 将项目版本提升至 0.3.02，配置版本提升至 0.3.02
• 更新依赖项：openai 至 1.87.0，vllm 至 0.10.0，torch 至 2.7.1，添加 torchvision，transformers 至 4.53.2，以及 triton 至 3.3.1

CI：

• 将 pre-commit-hooks 升级至 v6.0.0，ruff 升级至 v0.12.8