Skip to content

TeXada Desktop v0.4.2-beta.4

Pre-release
Pre-release

Choose a tag to compare

View all tags
@github-actions github-actions released this 16 Oct 02:17
v0.4.2-beta.4
1329999
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

TeXada Desktop 0.4.2-beta.4

中文

  • 修复 MiniCPM-V 4.7 1B 通过 Ollama 连接时关闭思考参数无效的问题：文本、规划、模型补全和 OCR 使用 reasoning_effort=none。其他兼容服务保留原有适配。
  • 内置 llama.cpp 更新到 b11669，固定 macOS arm64/x64 与 Windows CPU 安装资源的 SHA256。Ollama 模型已注册与实际推理可用分开展示，不再将模型列表成功误报为推理就绪。
  • 整合文本语义保护：显式集合运算组合、上下标关系、数列极限、普通高阶导数、曲线闭合性、矩阵乘积与行列式、阶乘、绑定操作数的向量二范数；不以任意占位符替代缺失对象。
  • M4 / 16GB 上 TeXada + Ollama b11669 修复后实测：100 条自然语言请求均返回 HTTP 200，92 条 API 工作流 valid=true，53 条严格结构匹配；8 张合成 OCR 图片均成功，7 张结构匹配。空公式模型失败由 4 次降为 0。6 条补全全部走确定性规则，不证明模型补全能力。结果详见 docs/ollama-b11669-validation.md。
  • 仍需人工核对： 结构匹配不是数学正确率；26 条已提交公式与严格参考不一致，尚无独立人工评审，也未验证手写或真实页面。保留人工审查机制。本版继续作为 Beta，不替代 v0.4.1 稳定版。

English

  • Use Ollama's reasoning_effort=none for MiniCPM-V 4.7 1B text, planning, model completion and OCR, instead of ignored template kwargs; preserve other provider profiles.
  • Pin bundled llama.cpp to b11669 with verified platform asset hashes. Distinguish Ollama registration from observed inference readiness.
  • Integrate text guards for explicit paired set operators, script roles, sequence limits, ordinary derivative order, curve closure, symbolic matrices, factorial notation and operand-bound vector norms.
  • Local M4/16GB validation: NL HTTP 200 100/100, syntax-valid 92/100, strict structure MATCH 53/100; synthetic OCR 8/8 successful and 7/8 MATCH. Empty candidate failures fell from four to zero. Six completion cases were deterministic, not model-quality evidence.
  • Structural match is not mathematical correctness; 26 committed formulas differ from strict references. No independent human adjudication or handwriting/real-page acceptance. Keep this release as a Beta; stable v0.4.1 remains latest.

Installers: signed and notarized macOS arm64/x64 DMGs and Windows x64 NSIS.
Author: CacinieP

Installed M4 package validation / M4 安装包实测

  • Mounted the actual signed/notarized arm64 DMG and launched its packaged shell, backend and b11669 runtime. Installer SHA256 matched the release asset. CI passed 1064 Python tests (13 skipped), 24 frontend tests, three platform builds and native installation smoke.
  • Packaged built-in path: 100/100 NL HTTP 200; API workflow valid=true 90/100; strict structure MATCH 50/100. Synthetic OCR: 8/8 HTTP 200, workflow valid=true 7/8, structure MATCH 7/8. Six completion examples were deterministic. These are separate from the source Ollama before/after measurements above.
  • Installed Ollama path: six text and two OCR requests returned nonempty outputs; all recorded calls used reasoning_effort=none. Text/OCR readiness changed from unverified to verified after actual inference.
  • One integral OCR candidate compiled and matched its reference, but an unsupported export format proposed by the model was rejected and retained for review; this repeated during diagnosis. Twenty-eight committed NL formulas differ from strict references. No claim of mathematical correctness, independent human adjudication, real-page/handwriting OCR, or Intel/Windows live-model accuracy.
  • 本机已验证实际 DMG 安装、签名、公证及打包后端。文本、OCR 均可调用；HTTP、工作流 valid、结构匹配与数学正确性分别统计。仍有模型错误和被拦截的工具调用，保留人工审查与 Beta 定位。

Detailed installed validation: https://github.com/CacinieP/TeXada-the-Math-Agent/blob/main/docs/beta4-installed-validation.md

Assets 7
Loading