1
00:00:00,280 --> 00:00:05,420
Imagine a model choosing its next word, with several
continuations that fit the same context.
想象模型选择下一个词，同一上下文中有几个合理的续写。

2
00:00:06,160 --> 00:00:11,340
Our first bar is the observed label; the others are alternatives
in this original teaching example.
第一根柱表示观测标签，其余柱表示这个原创示例中的备选项。

3
00:00:12,340 --> 00:00:18,600
A training example records one continuation, without proving
that every other plausible continuation is wrong.
样本记录一种续写，但这并不证明其他续写全部错误。

4
00:00:19,560 --> 00:00:25,440
Imagine an update pushing nearly all probability onto the label,
making the alternatives difficult to sample.
现在想象更新把几乎全部概率推给标签，使其他续写难以采样。

5
00:00:26,400 --> 00:00:32,500
These probabilities are illustrative, but increasing target
confidence can also erase other useful possibilities.
这些概率仅供示意，但增强目标信心也可能抹去其他有用可能性。

6
00:00:33,520 --> 00:00:40,080
Can supervised adaptation learn the observed answer while
preserving useful relationships among the other possibilities?
问题是监督适应能否学习观测回答，同时保留其他可能性之间的有用关系。

7
00:00:40,900 --> 00:00:47,620
The paper studies pretrained distributions, rather than treating
every non-label token as uninformative tail noise.
论文研究预训练分布，而不是把每个非标签词元都视为无用尾部噪声。

8
00:00:48,820 --> 00:00:54,380
Entropy analysis finds comparable candidate probabilities,
corresponding to different effective support sizes.
熵分析发现概率大致相当的后续词元群，对应不同有效支持集大小。

9
00:00:55,560 --> 00:01:01,560
Shannon and Renyi entropies describe effective support size and
how close the normalized profile is to uniform.
香农熵与 Rényi 熵描述支持集大小及归一化分布接近均匀的程度。

10
00:01:02,680 --> 00:01:08,880
Small groups can reflect semantic branching; larger groups can
reflect structural ambiguity among plausible digits.
小群可能表示语义分支，大群可能表示多个合理数字等结构歧义。

11
00:01:10,240 --> 00:01:16,640
Reported patterns span instruction, code, and mathematics
datasets across different model families and architectures.
报告模式跨越通用指令、代码和数学数据，以及不同模型家族、架构与规模。

12
00:01:17,840 --> 00:01:22,940
The goal preserves plausible alternative structure, while
allowing their absolute probabilities to change.
应保留的是合理备选项的组织方式，而非要求每个绝对概率固定。

13
00:01:23,767 --> 00:01:31,027
LP-SFT selects local candidates from the frozen base model,
using its highest ranked next-token possibilities.
LP-SFT 首先从冻结基础模型的高排名后续词元中选择局部候选集。

14
00:01:31,967 --> 00:01:38,307
It removes the supervised label, leaving only non-label
candidates as the explicit objects of preservation.
它移除监督词元，把非标签候选项作为明确的保留对象。

15
00:01:39,307 --> 00:01:45,247
Cross-entropy pushes the label upward; preserving that same
label could directly oppose its supervision.
交叉熵已提高标签；保留同一标签可能在其相对偏好增强时对抗监督。

16
00:01:46,167 --> 00:01:51,827
The target learns from the example, while the preservation term
protects relationships among the alternatives.
分离明确了两项职责：目标学习样本，备选项保留相互关系。

17
00:01:52,707 --> 00:01:59,507
The default uses ten base candidates, covering several plausible
regimes without a separate per-token schedule.
默认使用十个基础候选项，覆盖多种合理状态，无需逐词元安排候选预算。

18
00:02:00,287 --> 00:02:05,567
This is a training objective, rather than an inference-time
requirement to filter generated words.
这是训练目标设计，并不要求部署模型用这些候选项过滤生成词。

19
00:02:06,233 --> 00:02:12,033
Return to the toy distribution, with an update raising the label
while retaining alternative proportions.
回到示意分布：另一种更新提高标签，同时让备选项保持比例。

20
00:02:13,013 --> 00:02:20,073
The alternatives start at thirty, twenty, and ten percent:
relative proportions of three to two to one.
最初备选项占百分之三十、二十、十，相对比例为三比二比一。

21
00:02:20,833 --> 00:02:27,473
Their updated probabilities are fifteen, ten, and five percent,
so their combined mass falls by half.
示意更新后概率为百分之十五、十、五，因此总概率减半。

22
00:02:28,513 --> 00:02:35,513
Inside the non-label set, the relative profile remains;
meanwhile, the supervised label gains probability.
非标签集内部仍保持相同比例，而监督标签获得更多概率。

23
00:02:36,553 --> 00:02:42,473
Local normalization compares the alternative set shape,
independently of the total probability assigned to it.
局部归一化明确这种区别：独立于总概率，比较备选集内部形状。

24
00:02:43,513 --> 00:02:50,253
These ratios illustrate the goal; real training encourages
preservation through a loss, rather than invented values.
精确比例展示意图；实际训练用损失鼓励匹配，而非强制这些虚构值。

25
00:02:50,860 --> 00:02:57,780
The objective combines target cross-entropy with KL over locally
normalized non-label candidate distributions.
完整目标结合目标交叉熵与局部归一化非标签分布上的 KL 损失。

26
00:02:58,880 --> 00:03:03,960
Cross-entropy learns observed response tokens from supervised
instruction and response examples.
交叉熵提供适应信号，学习微调指令与回答样本中的观测响应词元。

27
00:03:04,940 --> 00:03:10,520
The second term uses fixed base information to protect relative
preferences among selected alternatives.
第二项比较当前可训练分布与固定基础信息，保护所选备选项的相对偏好。

28
00:03:11,720 --> 00:03:18,260
Gradient analysis explains why unnormalized matching instead
pushes more probability mass toward the alternatives.
梯度分析表明，未归一化匹配会把更多概率推向备选项。

29
00:03:19,380 --> 00:03:25,620
Local normalization removes that pressure; target exclusion
prevents direct conflict with the label supervision.
局部归一化移除该压力；排除目标避免直接对抗标签自身监督信号。

30
00:03:26,580 --> 00:03:31,720
The frozen base is a local structural reference, rather than a
stronger teacher producing new answers.
冻结基础模型因此是局部结构参考，而非生成新回答的更强教师。

31
00:03:32,347 --> 00:03:37,687
A naive implementation would keep both models in memory and run
the frozen base at each step.
直接实现会把两模型放在内存中，每一步同时计算冻结基础模型与可训练模型。

32
00:03:38,427 --> 00:03:44,847
LP-SFT uses the fact that its base model and training examples
are fixed before optimization begins.
LP-SFT 利用基础模型和监督训练样本在优化前已经固定这一事实。

33
00:03:45,507 --> 00:03:51,947
An offline pass computes base candidate logits and caches the
information needed for local preservation sets.
一次离线计算生成基础候选 logits，并缓存构造局部保留集所需信息。

34
00:03:52,827 --> 00:04:00,027
Fine-tuning compares current model logits with that cache,
restricting the comparison to relevant non-label candidates.
微调结合当前模型 logits 与缓存，只在相关非标签候选集上比较。

35
00:04:00,727 --> 00:04:06,687
This avoids an extra frozen-model forward pass at each step
while retaining the local base preferences.
这移除每一步的额外冻结模型计算，同时保留基础模型局部偏好。

36
00:04:07,707 --> 00:04:13,887
Cache construction remains a separate, one-time preparation cost
alongside the ongoing fine-tuning computation.
缓存仍有一次性准备成本，与报告微调配置的持续训练计算分开。

37
00:04:14,507 --> 00:04:18,847
Table one uses Qwen version three, with fourteen billion
parameters.
表一使用 Qwen3-14B，即参数规模为一百四十亿的 Qwen 第三版模型。

38
00:04:19,667 --> 00:04:24,067
It is fine-tuned on mixed mathematics, code, and general
instruction data.
它在数学、代码和通用指令混合数据上进行微调。

39
00:04:24,947 --> 00:04:29,987
The aggregate averages displayed single-sample, finite-budget,
and general knowledge metrics.
总体值是所有展示指标均值，结合单次、多次采样测试和通识准确率。

40
00:04:30,847 --> 00:04:37,467
The base scores sixty-six point seven nine; cross-entropy
fine-tuning reaches fifty-eight point five five.
此混合设置中，基础模型为六十六点七九，普通交叉熵微调为五十八点五五。

41
00:04:38,287 --> 00:04:45,867
LP-SFT reaches sixty-five point three two, reducing the gap
while remaining below this configuration base average.
LP-SFT 达到六十五点三二，缩小差距，但此配置下仍低于基础均值。

42
00:04:46,926 --> 00:04:54,327
MMLU accuracy likewise stays closer to the base under LP-SFT
than cross-entropy in this matched comparison.
在这一匹配比较中，LP-SFT 的 MMLU 准确率也比交叉熵更接近基础模型。

43
00:04:55,127 --> 00:05:01,167
These scoped retention results use a mean of displayed metrics,
rather than pooled benchmark task accuracy.
这些是限定的保留结果；总体值是指标平均，而不是汇总任务准确率。

44
00:05:01,773 --> 00:05:08,093
Separate the supervised lesson from the alternative structure
that the pretrained model already considers plausible.
核心思想是分开样本教授的内容与预训练模型已有的备选结构。

45
00:05:09,073 --> 00:05:16,733
Target exclusion protects supervision; local normalization
preserves relationships without fixing alternative mass.
移除目标保护直接监督；局部归一化保留关系，而不冻结备选项总概率。

46
00:05:17,873 --> 00:05:23,953
These choices strengthen the label while encouraging non-label
preferences to survive domain adaptation.
两者允许标签增强，同时鼓励有用的非标签偏好在领域适应中保留。

47
00:05:25,153 --> 00:05:31,153
Evaluation tests one sampled answer and access to correct
solutions within a finite sampling budget.
评估同时考察单次回答，以及有限采样预算内是否仍能找到正确解。

48
00:05:32,153 --> 00:05:38,353
Matched results support preservation, while the objective
specifies which relationships receive protection.
匹配比较支持这种保留目标，精确训练目标则说明它试图保留何种结构。

49
00:05:39,353 --> 00:05:47,093
As the observed answer improves, LP-SFT asks which other
plausible continuations should remain possible after training.
LP-SFT 为监督学习新增问题：观测回答改善时，哪些合理路径应当仍然可能？