首页 / 热度
热度
GitHub 元数据取自 2026-11-04。
本周上升
Stars gained since the oldest snapshot of the last seven days.
- 01jev-ultrafast一次请求同时从索引化的 DOM 表中选出操作和目标元素;小模型只负责写带类型的文本。苏黎世到伦敦的订票用时 7.1 秒。
- 02jevgrep by dzhng向 Jev 询问哪些声明能回答关于仓库的自然语言问题,从而为编码代理找出所需的文件、声明和原文摘录。
- 03kevQwen2.5-0.5B 上的 LoRA 适配器与读出头,一次 prefill 回答多个带类型问题;在 MacBook 上两小时内训完,留出集 ECE 0.065,说 TypeSafe 的传输格式。
- 04eveVercel 的智能体框架;Jev 在其 evaluate 步骤中充当带类型的评判。
- 05AnyJev把任意开源 Hugging Face 模型变成通过 vLLM 提供服务的校准决策端点,前两个级别无需微调。
- 06hermes-jev-skills把智能体的小决策交给 Jev:这一轮由哪个模型回答、加载哪些技能、哪些检索片段重要、哪些对话轮次能留到压缩之后。
- 07ai-cli终端里的 Vercel AI SDK,其 evaluate 路径跑在 Jev 上。
- 08fast-jev-compaction用 Jev 的判定替代上下文压缩的摘要:一次请求为每个工具调用及其结果打分,丢弃过时项,保留的一律原样留存。
- 09skills用于设计问题、构建工作流并做评估的智能体技能。
- 10jev-seo by AgriciDaniel爬取网站,对照 52 条 SEO 规则检查,由 Jev 评判每个页面,并生成 PDF、电子表格和 Markdown 报告。
- 11pg-jevPostgreSQL 扩展,用自然语言回答关于你的表的问题。
- 12SemIf在单张 3090 上用开源模型实现语义 if;星标最多的独立复刻实现,前身为 openjev。
星标最多
Every GitHub-hosted project on the list, by star count.
- 01jev-ultrafast一次请求同时从索引化的 DOM 表中选出操作和目标元素;小模型只负责写带类型的文本。苏黎世到伦敦的订票用时 7.1 秒。
- 02kevQwen2.5-0.5B 上的 LoRA 适配器与读出头,一次 prefill 回答多个带类型问题;在 MacBook 上两小时内训完,留出集 ECE 0.065,说 TypeSafe 的传输格式。
- 03fast-jev-compaction用 Jev 的判定替代上下文压缩的摘要:一次请求为每个工具调用及其结果打分,丢弃过时项,保留的一律原样留存。
- 04eveVercel 的智能体框架;Jev 在其 evaluate 步骤中充当带类型的评判。
- 05SemIf在单张 3090 上用开源模型实现语义 if;星标最多的独立复刻实现,前身为 openjev。
- 06jevgrep by dzhng向 Jev 询问哪些声明能回答关于仓库的自然语言问题,从而为编码代理找出所需的文件、声明和原文摘录。
- 07jev-trader在 Monad 的每个区块做一次交易决策,标的为 Kuru MON-USDC,每次约 300 ms。
- 08skills用于设计问题、构建工作流并做评估的智能体技能。
- 09NanoJev0.6B 的复刻实现,支持并行决策、动态候选项和端到端训练流水线。
- 10pi-mcp-adapter对 MCP 工具结果做可选的带类型评估和语义搜索,受按服务器配置的数据外发白名单约束。
- 11AnyJev把任意开源 Hugging Face 模型变成通过 vLLM 提供服务的校准决策端点,前两个级别无需微调。
- 12ollaya拉取并运行 Laya、kev、JevK5 等开源决策模型,通过兼容 TypeSafe 的本地端点提供服务,就像 Ollama 服务 LLM 一样。
最近提交
Last commit to the default branch, as GitHub reports it.
- 01quackd面向 LLM 驾驶机器人的命令行,覆盖七种本体,带可选的 Jev 步进器,只在若干调用之间做选择,从不直接写关节角。
- 02eveVercel 的智能体框架;Jev 在其 evaluate 步骤中充当带类型的评判。
- 03cultivarPinecone 的技能测试 CLI,带 Jev 评分后端,据报比 LLM 评分器便宜约 30 倍。
- 04neurolinkTypeScript AI SDK,其中通过 Jev 实现的 decide 与 generate、stream 并列:一次类型化判断调用即可路由模型选择、裁剪上下文并挑选 MCP 工具。
- 05OneJevOpen multimodal System One model, 0.8B to 27B, for screenshots, photos, video and text.
- 06muCoding agent and desktop app built on Pi that hands 38 routine decisions to Jev, from which tool output enters the context to prompt-injection screening and checking a "done", each one switchable between active, shadow and off with every verdict in a local ledger; pre-release.
- 07typesafeai-dotnet-sdk社区版 .NET SDK,支持带类型的 Noul、Choice 和 Score 问题。
- 08Valen基于 Qwen3.5 的多模态决策模型,除文本外还能对照图像和视频为候选打分,附训练代码和开放权重。
- 09pi-mcp-adapter对 MCP 工具结果做可选的带类型评估和语义搜索,受按服务器配置的数据外发白名单约束。
- 10pi-warden引导而非打断的护栏:不可逆调用、偏离任务的调用、卡住的循环、未经验证的完成声明,每次约 250 毫秒。
- 11trade-jev在 NQ 订单簿数据上回测把 Jev 当作买入、卖出或持有的交易者。
- 12ai-cli终端里的 Vercel AI SDK,其 evaluate 路径跑在 Jev 上。
本月新增
First time each entry appeared in the readme.
- 01muCoding agent and desktop app built on Pi that hands 38 routine decisions to Jev, from which tool output enters the context to prompt-injection screening and checking a "done", each one switchable between active, shadow and off with every verdict in a local ledger; pre-release.
- 02OneJevOpen multimodal System One model, 0.8B to 27B, for screenshots, photos, video and text.
- 03building-with-typesafe-jevSkill that packs the Choice, Score, and Noul API shapes, confidence thresholds, and 150+ community projects with a code sketch each, so an agent can design without fetching the docs.
- 04jev-skill-routerHook that asks Jev which installed skill fits each prompt and logs the answer; in the author's week of shadow mode, 28 of 539 suggestions were followed by a call of that skill, and the author then removed it.
- 05Jev RAGSearches local documents with BM25 or hybrid retrieval, reranks candidate passages with batched Jev Noul judgments, and measures seven retrieval pipelines on BEIR NFCorpus.
- 06WaterSheepOpen-weight ModernBERT-base fine-tune that answers Noul, Choice, Score, and multi-label questions with a probability per option, serves `/v1/systemone` for the official Python SDK, and runs in the browser through ONNX; independent replica, not TypeSafe weights.
- 07JevSpanZero-shot named entity recognition for Chinese and English: code proposes candidate spans at punctuation and Jev choice questions nominate, verify, and fix the boundary of each entity; reports 73.7 average strict F1 on 12 NER benchmarks against 72.1 for Qwen3.8-27B.
- 081 Million EmojisShared 1000 × 1000 emoji canvas where Jev joins each stroke with one Choice over named emoji and square pairs, its pick sampled from the probabilities; [paint on it](https://chriswijnia.com/lab/emoji).
- 09Jev in the WildSurveys 2,170 public Jev projects and maps early growth, application domains, and how Choice, Noul, and Score support different decisions.