← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 6501.k · 生成模型:GPT-6

外部检查点需绑定训练条件

理解等级:中级

下载的作者检查点可能采用不同 rollout 数或预算;即使架构兼容,也不等于本地示例配置产生的相同结果。比较时记录文件哈希、来源与训练设置,再用固定评估协议验收,不把书中成绩写成本机复现。

这张卡片看不懂?

已有反馈 · 0 条

    还没有人提交反馈。

    在整书笔记中的位置

    原书出处 · 1 处

    • · 查看原书依据

      available for evaluation or further experimentation: from reasoning_from_scratch.qwen3 import download_qwen3_grpo_checkpoints download_qwen3_grpo_checkpoints(grpo_type="no_kl", step="00050") Note that this checkpoint was created with similar settings as listing 6.17 except that num_rollouts=4 was increased to num_rollouts=8. These checkpoints are fully compatible with the model evaluation utilities introduced in chapter 3, which allows us to evaluate RL-trained models using the same MATH-500 verification pipeline. For convenience, you can reuse the evaluation

    参考来源

    - 位置:《Build a Reasoning Model (From Scratch)》第06章《强化学习训练推理模型》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Build a Reasoning Model (From Scratch)/_epub-src/OEBPS/Text/chapter-6.html) - 结论依据:摘录中关于外部检查点需绑定训练条件的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。 - 原始内容:available for evaluation or further experimentation: from reasoning_from_scratch.qwen3 import download_qwen3_grpo_checkpoints download_qwen3_grpo_checkpoints(grpo_type="no_kl", step="00050") Note that this checkpoint was created with similar settings as listing 6.17 except that num_rollouts=4 was increased to num_rollouts=8. These checkpoints are fully compatible with the model evaluation utilities introduced in chapter 3, which allows us to evaluate RL-trained models using the same MATH-500 verification pipeline. For convenience, you can reuse the evaluation