Strands Decider 2B：一个小的开源决策模型，只做选择题和打分 Strands Decider 2B: a small open decision model that picks an option or gives a score

来源Sources

为什么推荐Why we recommend it 博文命令示例仍写 v19。规格和分数以 README 为准，博文里的名次不转抄。本库未 pip install，未下载权重，未复跑 JevBench，未在 RTX 3090 或 Mac 上计时。同板与 Clef、Gutsy、Jeeves、Jev、Laya、Ollaya 互链；Cases 里已有本机 Jeff 和命令行 Onesie。 The blog’s command examples still say v19. Specs and scores follow the README. This library did not pip install, did not download weights, did not re-run JevBench, and did not time it on an RTX 3090 or a Mac. Cross-linked on this board with Clef, Gutsy, Jeeves, Jev, Laya, and Ollaya. Cases already lists local Jeff and the Onesie CLI.

要点（官方称 / 社区称）Highlights (vendor / community claims)

• 它是决策模型：不自由写字。你给现状，再给允许的答案。问题有三种，仓库用 noul（是或否）、choice（多选一）和 score（按分数尺打分）。一次前向就把这些答案的概率算出来。仓库写参数量约 19 亿，躯干是 Qwen3.5-2B-Base，原来的语言模型头被拿掉，换成约一百万参数的指针头，再用 rank-16 的 LoRA 微调。LoRA 是只训练一小部分新增权重的做法。 It is a decision model: it does not write free text. You provide a situation and the answers you allow. The repo’s three question types are noul (yes or no), choice (one of several), and score (a point on a scale). One forward pass produces the probabilities. The repo says about 1.9 billion parameters. The torso is Qwen3.5-2B-Base, the old language-model head is removed, a pointer head of about a million parameters replaces it, and a rank-16 LoRA adapts the torso. LoRA means only a small set of added weights is trained.
• 怎么试（仓库写的命令，本库没跑过）：pip install strands-decider，然后 strands-decider ask，模型名写 StrandsAgents/strands-decider-2B-hobson-v21。也可以在本机 serve。仓库写它能在 CPU、苹果芯片 Mac 或一块消费级显卡上跑。许可 Apache-2.0。 How to try, as the repo writes it (not run here): pip install strands-decider, then strands-decider ask with the model id StrandsAgents/strands-decider-2B-hobson-v21. You can also serve it locally. The repo says it runs on CPU, an Apple-silicon Mac, or a consumer GPU. License Apache-2.0.
• 仓库自测，本库未复跑，也不把它当成排名。README 的表：JevBench 公开集 231 题，v21 准确率 0.762（176/231），v19 为 0.723（167/231）。同一张表写，六次种子的平均和 v19 配方几乎一样，单次相差大约 10 题以内就算没有分出胜负。延迟列在 v19：RTX 3090 中位 115 ms。官博 10 月 1 日还有一套“在 2B 档第几名”的说法，本卡不转述，因为仓库后来的表自己提醒小差距不可靠。 Repo self-test, not re-run here, and not treated as a ranking. The README table: on the 231-task public JevBench set, v21 accuracy is 0.762 (176/231) and v19 is 0.723 (167/231). The same table says a six-seed average of the v21 recipe and the v19 recipe land in the same place, and a gap under about 10 tasks between two single runs is unresolved. The latency column is on v19: median 115 ms on an RTX 3090. The October 1 blog also has a “place in the 2B class” claim. This card does not repeat it, because the repo’s later table says small gaps are not reliable.