Clef 上线 Ollama 云端

Clef 基于 Qwen3.8-27B 微调而来,并以 Apache 2.0 许可证发布。
多模态:可基于文本、JSON 或图像(包括截图、收据、表单和照片)做出决策。
单次通过:Clef 只读取输入一次,同时为每个问题的每个答案打分,而不是逐词写出回复。
领先的准确率:Clef 位居 Cloudflare 的 Decision Index 榜首,并在其基准测试套件的大多数项目中胜过 Jev。
Clef 支持以下使用场景:
路由请求:定义目标以及每个目标的适用条件。Clef 返回所选目标以及每个目标的概率。
检查条件:定义一个是或否的问题以及证据。Clef 返回 true 或 false,以及各自的概率。
应用策略:定义规则和允许的结果,即可根据提供的文本或图像获得带类型的决策。
读取文档:提供页面照片以及需要评估的字段,以带概率的方式逐字段做出判断。
工具调用审核:在智能体的工具调用执行前对其进行检查。
来自 Cloudflare 公告的结果,依据 Jev Decision Index 的评估进行打分:
家用电器 · 精确匹配(区分大小写)
在 Typesafe 自身的工作流评测中,Clef 在 4 个领域中 3 个胜过 Jev:
智能体追踪可观测性
43 项评测基准上的延迟,单位为毫秒:
更多信息,请参见:
译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(Clef is available on Ollama's cloud)
Clef is fine-tuned from Qwen3.8-27B and released under the Apache 2.0 license.
Multimodal: Make decisions over text, JSON, or images, including screenshots, receipts, forms, and photos.
One pass: Clef reads the input once and scores every answer to every question at the same time, instead of writing a response word by word.
Leading accuracy: Clef tops Cloudflare’s Decision Index and beats Jev on most of its benchmark suite.
Clef supports the following use cases:
Route a request: Define destinations and when each applies. Clef returns the chosen destination and each destination’s probability.
Check a condition: Define a yes-or-no question and the evidence. Clef returns true or false and the probability of each.
Apply a policy: Define rules and allowed outcomes to receive a typed decision based on supplied text or images.
Read a document: Supply a photo of a page and the fields to evaluate for field-by-field decisions with probabilities.
Tool-call moderation: Check an agent’s tool call before it runs.
Results from Cloudflare’s announcement, scored on evaluations from the Jev Decision Index:
Home appliances · case exact
On Typesafe’s own workflow evals, Clef beats Jev in 3 of 4 areas:
Agent trace observability
Latency across 43 eval benchmarks, in milliseconds:
For more information, see:
原文为英文,正文由自动抓取并保留原文。
原文发布于 Ollama Blog(Clef is available on Ollama's cloud),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。