Introducing Clef: our open-source decision models, and new RL fine-tuning platform
- ID: 317971da
- 原文链接: https://blog.cloudflare.com/clef-decision-models
- PDF: N/A
- 作者: Cloudflare Blog
- 日期: 2026-10-01
- 更新: N/A
- 分类: models
- 来源类型: article
- 标签: decision-models, jev, classification, inference
- 质量评分: 4/5
- 抓取时间: 2026-10-02T12:30:30.134235+00:00
中文导读
跟进 Typesafe 的 Jev 决策模型概念，Cloudflare 发布开源权重（Apache 2.0）决策模型 Clef 与 Clef-flash，托管在 Workers AI 且与 Jev API 兼容，在 Jev Decision Index 上居首与 Jev 差异：带视觉编码器可分类图像（Jev 仅文本）64k 上下文（Jev 32k）架构上冻结 Qwen3.8-27B / Qwen3.5-9B backbone，加 rank-256 LoRA 与两阶段 attention routing，推理时 prefill-only 一次前向并行给所有合法 schema 选项打分决策步非自回归无逐 token 生成，中位延迟 209.3ms/38.8ms（Jev 524.1ms）；域名分类示例 2.2s 完成（gpt-oss-120b 4.7s）训练用 label-smoothed CE + Brier 校准 + 自研 RLCD，同时开出 RL 微调产品线
为什么值得关注
Cloudflare 开源决策模型 Clef：冻结 Qwen + prefill-only 并行打分，比 Jev 快且带视觉，Jev API 兼容
关键信息
- 原文标题: Introducing Clef: our open-source decision models, and new RL fine-tuning platform
- 作者: Cloudflare Blog
- 原文链接: https://blog.cloudflare.com/clef-decision-models
- 发布时间: 2026-10-01
- 关联标签: decision-models, jev, classification, inference
English Excerpt
"During inference, Clef uses Qwen for a prefill-only pass, then scores the valid schema choices in parallel. The decision step is non-autoregressive, so there's no intermediate text to generate token by token, making Clef significantly faster than autoregressive LLMs."
"First, it has a vision encoder so it's able to take in images and classify visual content. This is different from Jev, which only does text classification today. Secondly, our model has a 64k context window (compared to Jev's 32k)."
"By freezing Qwen3.8-27B for Clef and Qwen3.5-9B for Clef-flash, we jointly optimized the routing head alongside rank-256 low-rank adapters. Our post-training utilizes label-smoothed cross-entropy for valid schema outputs paired with a Brier loss to refine probability calibration. [...] We also developed Reinforcement Learning for Calibrated Decisions (RLCD) to serve as a secondary optimization target, granting partial credit to adjacent ordinal choices, rewarding fully precise record outputs, and applying a reference penalty to prevent distribution shift."
English Summary
Following Typesafe's Jev decision-model concept, Cloudflare releases open-weight (Apache 2.0) decision models Clef and Clef-flash hosted on Workers AI, Jev-API compatible and topping the Jev Decision Index. Differences vs Jev: a vision encoder for image classification and a 64k context window (Jev: text-only, 32k). Architecture: frozen Qwen3.8-27B / Qwen3.5-9B backbones plus rank-256 LoRA and two-stage attention routing; a prefill-only pass scores all valid schema choices in parallel a non-autoregressive decision step with 209.3ms/38.8ms median latency vs Jev's 524.1ms (domain classification demo: 2.2s vs gpt-oss-120b's 4.7s). Training uses label-smoothed cross-entropy, Brier calibration, and a novel RLCD; an RL fine-tuning product accompanies the launch.
Obsidian Notes
- Content grounded in an opencli fetch during content-fetcher backfill (papers via
opencli arxiv paper -f json; articles viaopencli web read). - 中文导读 block is the entry's summary_zh (verbatim); the value note is the entry's one_liner; English blocks are verbatim excerpts from the fetched source. No claims beyond the fetched metadata/source text were added.