RCreddit.com
25
·开发者社区 · RSS
ThinkingCap-Qwen3.6-27B: same accuracy as base Qwen3.6 with ~50% fewer thinking
这条记录涉及编程工具或代码能力更新，适合开发者评估工作流变化和可复用价值。
ThinkingCap-Qwen3.6-27B 在通用推理、非推理选择题问答、日常多轮对话、系统提示遵循、安全性、数学、代码和代理用例等多个方面进行了严格评估。由于 Qwen 推荐的采样温度 1.0 下推理质量存在高度可变性，研究人员通过多个随机种子运行每个基准测试，并对所有结果进行统计显著性检验。…
We rigorously evaluate the resulting checkpoint across general reasoning, non-reasoning multiple-choice question answering, everyday multi-turn conversations, system prompt adherence, safety, math, code and agentic use cases. Due to the high variability of reasoning quality at Qwen-recommended sampling temperature 1.0, we run each benchmark with multiple seeds and do statistical significance testing on all the results. We evaluate both in domain (holdout parts of selected datasets included in training) and out of domain.
主题标签Qwen模型发布开源代码
原始关键词#thinkingcap#accuracy#thinking#fewer#qwen3#base