仓库/Skill 库

🔍 使用多种 prompt 技巧在多步数学问题上分析 Mistral-7B 模型的数学推理能力。🔍 Analyze the mathematical reasoning abilities of the Mistral-7B model using diverse prompting techniques on multi-step math problems.

与供应商无关的 coding-agent 框架，具备确定性工作流图、持久化证据与 fail-closed 沙箱执行。Provider-neutral coding-agent harness with deterministic workflow graphs, durable evidence, and fail-closed sandboxed execution

在小型 LLM 中植入认知美德的激活引导实验——三项发现与一个失败模式数据集（DOI 见 README）Activation-steering experiments on installing epistemic virtues in small LLMs — three findings + a failure-mode dataset (DOIs in README).

在 DABench 数据分析任务上对 DSPy RLM（Recursive Language Models）进行基准测试，使用自动评分实现基于代码的迭代评估Benchmark DSPy Recursive Language Models on DABench data analysis tasks with automated scoring for iterative code-based evaluation

MeterSphere v2.10 个人开发分支 — 一站式开源持续测试平台。工作流引擎(Flowable 7) · 需求池(从0到1) · 微前端(qiankun→micro-app) · AI知识库 · 测试跟踪，含个人笔记与创意项目。

HateMirage——可解释的伪仇恨检测与多维推理，ICON 2026 共享任务。在 HateMirage 语料库（4,530 条带标注的伪仇恨评论）上进行目标识别 + 意图与隐含意义生成。HateMirage - Explainable Faux Hate Detection and Multi-Dimensional Reasoning - Shared Task @ ICON 2026. Target identification + Intent and Implication generation over the HateMirage corpus (4,530 annotated Faux Hate comments).

追踪并比较每一个严肃的 OpenClaw 替代方案——基于实测仓库数据与 AI 撰写的决策支持，且刻意保持独立。Track and compare every serious OpenClaw alternative — measured repo data, AI-written decision support, kept apart on purpose.

硕士论文（KCL）：在攻击者-机器留出划分下重新衡量 IoT 入侵检测，并测试基于 LLM 的报文预测是否能提升性能。移除泄露后，macro-F1 从 0.90 降至 0.60。MSc dissertation (KCL): re-measuring IoT intrusion detection under an attacker-machine holdout, and testing whether LLM-based packet prediction improves it. macro-F1 0.90 -> 0.60 once leakage is removed.

交互式仪表盘，对比五个 LLM（GPT-5.6 Luna、Mistral Small 4、DeepSeek v4 Flash、Gemma 4 31B 与 Qwen3.8 27B）对 12,349 篇西非法语区新闻文章的标注结果，包含评分者一致性、差异分析与盲法仲裁。法/英双语；此前的三轮三模型标注活动保留归档。Interactive dashboard comparing how five LLMs — GPT-5.6 Luna, Mistral Small 4, DeepSeek v4 Flash, Gemma 4 31B and Qwen3.8 27B — annotated 12,349 francophone West African press articles, with inter-rater agreement, discrepancy analysis and blind arbitration. Bilingual FR/EN; the earlier three-model campaign stays archived.

即时评估课程资格，输出明确的通过/未通过结果以及定制化的入学测试。Instantly evaluate course eligibility with clear pass/fail results and tailored entry assessments.

🎨 利用 AI 实时生成高保真 UI 设计，对比多版本方案，并跨多个模型导出可直接使用的代码。🎨 Generate high-fidelity UI designs in real-time with AI, compare variations, and export ready-to-use code across multiple models.

基于人类基准的 LLM 生成 Infrastructure-as-Code 安全基准测试。100 场景 × 12 模型配置 = 1,196 个工件，由 Checkov/Trivy/KICS 扫描，并与 634 个人工编写模板对比。所有模型的漏洞密度均为人工的 3.2–3.9 倍。arXiv:2608.28021Human-anchored security benchmark for LLM-generated Infrastructure-as-Code. 100 scenarios × 12 model configs = 1,196 artifacts scanned by Checkov/Trivy/KICS, compared against 634 human-written templates. Every model: 3.2–3.9× human vulnerability density. arXiv:2608.28021

语义回响实证数据仓库：多模型对照 + 全流程 7 模式 2026 评测结果 + 论文/图表