Skip to content

RimeJev 0.5.0 · Encrypted training capture

Choose a tag to compare

View all tags
@IIIIOvOIIII IIIIOvOIIII released this 11 Oct 15:46
· 3 commits to main since this release
v0.5.0
79e1d88

中文

RimeJev 0.5.0 增加独立、默认关闭的本机训练数据采集，并包含自 0.4.1 发布以来的同行语境补全改进。

  • 在“训练数据采集与管理…”中选择保存文件夹、仅采集或排除的应用、前文/模型字段及容量上限。初始为关闭、仅采集所选应用、名单为空；开启增强不会自动开启采集。
  • 关联真实拼音、有限前文、原生候选、展示顺序和实际人工选词。原始记录采用 AES-GCM 加密，密钥独立保存在本机钥匙串；采集不额外加载或调用模型。
  • 支持本机预览、移到废纸篓及离线 records/sft 导出。Tev prompt/completion 导出需用户提供兼容的本地分词器；导出是明文，应先审阅。数据管理先关闭采集，结束后保持关闭；不自动上传或训练。
  • 同行语境可使用已有、粘贴及停顿后继续输入的前文；空语境也可请求模型判断。继续保留原首选、按设置将不同推荐放在第二/三位；失败时原生输入继续可用。
  • 中英文 README 和 LLM_INSTALL.md 同步到固定 v0.5.0 安装入口。

验证覆盖采集开关和应用范围、异步加密存储、人工标签与模型结果分离、离线导出，以及现有真实 Rime 候选行为。专用合成验收应用实测了加密写入、预览、导出和可恢复删除，结束后采集关闭。Tev 指令格式仅通过合成分词器验证，未现场训练模型。

本次仅发布源码。源码和公开 Git 历史经过凭据及私有文件检查；不附带采集记录、训练导出、个人词库、凭据、配置、模型权重或本机二进制。仓库保留有明确说明的人工合成评估和测试样例。安装目标为 Apple Silicon / macOS 14+；没有 Apple 公证的即装安装包。模型推荐准确率和应用兼容性仍需分别验证。

English

RimeJev 0.5.0 adds independent, off-by-default local training capture and includes current-line context improvements since the 0.4.1 release.

  • Choose the capture directory, included or excluded apps, context/model fields and size limit. Defaults are off, selected-app scope and an empty list. Enabling enhancement never enables capture automatically.
  • Link actual pinyin, bounded context, native candidates, displayed order and observed human choices. AES-GCM encrypts raw records with a separate local Keychain key; capture does not load or call extra models.
  • Preview, recoverable deletion and offline records/sft exports are available. Tev prompt/completion export requires a compatible user-supplied local tokenizer. Exports are plaintext and need review. Data management disables capture and leaves it off; no automatic upload or training occurs.
  • Existing, pasted and paused same-line text can supply context; empty context can also request a model decision. The native first candidate remains stable, distinct recommendations use slots two/three, and native input remains available on failure.
  • Both README languages and the dedicated LLM installation guide now select the fixed v0.5.0 tag.

Validation covers opt-in/app scope, asynchronous encrypted storage, separation of human labels from model decisions, offline export and existing real-Rime candidate behavior. A dedicated synthetic app verified encrypted writes, preview, export and recoverable deletion; capture was left off. Tev instruction export was tested with synthetic tokenizers; no live model training was performed.

This is a source-only release. Source and public Git history were checked for credentials and private artifacts. No captures, training exports, personal dictionaries, credentials, local configuration, model weights or local binaries are attached. Documented hand-authored evaluation/test fixtures remain in the repository. The installation target is Apple Silicon / macOS 14+; an Apple-notarized installer is not provided. Model accuracy and application compatibility require separate validation.

Installation and verification / 安装与验证

The runtime is unchanged from tested and locally installed source 7cdde9e6ad3cc4375110d241b6bbfc3bec8a131b; publication changes only Markdown documentation. An existing installation of that 0.5.0 build does not need reinstalling for this source release. New installations must run preflight against their exact release checkout.

Assets 2
Loading