← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 6512.c · 生成模型:GPT-6

训练长度预算塑造可获奖轨迹

理解等级:中级

若答案在训练生成预算内未完成或未输出终值就拿不到奖励,预算会影响模型可探索、可获奖的轨迹。预算过小可能偏向短格式并排除长解法;更大预算提高资源成本,也不保证准确率继续增长。

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

      6.1 refers to the accuracy on the full 500-sample MATH-500 dataset we used in chapter 3. Note that the "Max tokens" column corresponds to the number of tokens that were allowed per rollout during training. If the number is 512, this encourages the LLM to provide the final boxed answer within this 512 token limit because otherwise it will not receive a reward during training.

    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:摘录中关于训练长度预算塑造可获奖轨迹的定义、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:6.1 refers to the accuracy on the full 500-sample MATH-500 dataset we used in chapter 3. Note that the "Max tokens" column corresponds to the number of tokens that were allowed per rollout during training. If the number is 512, this encourages the LLM to provide the final boxed answer within this 512 token limit because otherwise it will not receive a reward during training.