Model Checkpoint · Best Is Not Always Latest

01 · 先从日常问题开始 01 · Start with an everyday question

手机自动保存了五个版本，你会因为最后一个时间最新，就认定它一定最好吗？ Your phone saved five versions. Does the newest timestamp automatically make the last one best?

02 拖选检查点，对照验证损失和样例输出 Select a Checkpoint and Compare Validation Loss and Output

拖动检查点时间轴 drag checkpoint timeline

唯一主变量 One main variable

五个玩具检查点都能恢复，但验证损失先降后升。选择任一版本，面板回到当时参数、输出和成本状态。 All five toy checkpoints are recoverable, but validation loss falls and then rises. Select one to restore its parameters, output, and cost state.

检查点编号 Checkpoint number 3/5

固定种子 Fixed seed 20260722

教学模拟值，非真实模型输出 simulated for teaching — not real model output

验证损失 Validation loss

—

lower is better

样例正确率 Sample accuracy

—

fixed toy set

恢复成本 Resume cost

—

toy minutes

03 一个检查点至少保存四类状态 A Checkpoint Saves at Least Four Kinds of State

把刚才的变化拆开 Map the change to parts

01 · PARAM

参数 Parameters

保存当时模型数值，使该版本可被准确恢复。 Store model values so that exact version can be restored.

02 · OPT

优化状态 Optimizer state

若要继续训练，还需保存更新器进度等相关状态。 Continuing training also requires optimizer progress and related state.

03 · STEP

训练位置 Training position

记录处理到哪个批次，防止恢复后重复或跳过数据。 Record batch position to avoid repeating or skipping data after resume.

04 · EVAL

独立评估 Independent evaluation

最佳检查点由固定验证证据选择，不由时间戳决定。 Select the best checkpoint with fixed validation evidence, not timestamp.

04 最小选择规则：在固定验证集上取低损失 Minimal Selection Rule: Choose Lower Fixed-Set Loss

k* = arg min Lval(k)

k^* = arg min_k L_val(k)

五个玩具损失固定为 0.62、0.44、0.31、0.35、0.43，因此第 3 个优于最新的第 5 个。 Toy losses are fixed at 0.62, 0.44, 0.31, 0.35, and 0.43, so checkpoint 3 beats the latest checkpoint 5.

途中保存状态 Save during training 版本可恢复 version becomes recoverable

固定集评估 Evaluate on fixed set 比较同一量尺 use the same yardstick

标记最佳版本 Mark the best version 最佳不等于最新 best is not latest

05 边界实验：只选最新检查点 Boundary Test: Select the Latest Checkpoint Only

主动制造失败 Create a failure

第 5 个版本更新，却已从最佳点回退 Checkpoint 5 is newer but has regressed from the best

一键选择最新版本。验证损失从最佳 0.31 回升到 0.43，固定样例正确率也下降。 Select the latest snapshot. Validation loss rises from the best 0.31 to 0.43 and fixed-sample accuracy drops.

尚未执行。先在上方改变主变量。 Not run yet. Change the main variable above first.

✗ 最新检查点一定最好。 The latest checkpoint is always best. → ✓ 用固定验证证据选择，而不是看时间戳。 Select with fixed validation evidence, not timestamp.

✗ 只存模型文件就一定能续训。 A model file alone always resumes training. → ✓ 续训还需要优化状态和训练位置。 Resuming also needs optimizer state and training position.

✗ 检查点分数可跨不同测试条件直接比。 Checkpoint scores compare across different test conditions. → ✓ 比较必须使用同一数据与量尺。 Comparisons require the same data and yardstick.

06 · 一句话带走 06 · One line to keep

模型检查点保存训练途中的参数与相关状态，供恢复和评估；发布应选固定验证证据最好的版本，而不是机械选择最新版本。 A checkpoint saves parameters and related training state for recovery and evaluation. Release the version with the best fixed-set evidence, not automatically the newest.