1
00:00:00,240 --> 00:00:02,540
Imagine three records in a training mixture.
想象训练混合语料中的三条记录。

2
00:00:03,280 --> 00:00:05,060
One is a short mathematical derivation.
一条是简短数学推导。

3
00:00:05,740 --> 00:00:07,780
Another explains a software API.
一条解释软件 API。

4
00:00:08,500 --> 00:00:13,340
A third is an agent trace that calls a tool, sees an error, and
repairs its attempt.
第三条是智能体轨迹：调用工具、看到错误，再修正尝试。

5
00:00:14,340 --> 00:00:17,720
They may all be useful, but the evidence of usefulness looks
different.
它们都可能有用，但判断其价值的证据并不相同。

6
00:00:18,640 --> 00:00:20,300
MIRA starts from that selection problem.
MIRA 从这个数据选择问题出发。

7
00:00:21,060 --> 00:00:24,760
Mid-training mixes sources with different formats and capability
roles.
中间训练混合了格式和能力角色不同的来源。

8
00:00:25,460 --> 00:00:28,100
The question is not only how to score millions of records
quickly.
问题不只是如何快速给数百万条记录评分。

9
00:00:28,600 --> 00:00:34,280
It is what should count as quality inside each source group,
before those scores decide what the model will see.
还包括在分数决定模型所见内容之前，各组内部应该如何定义质量。

10
00:00:35,127 --> 00:00:38,547
The experiments organize twenty-one sources into five groups:
实验把二十一个来源组织成五组：

11
00:00:39,187 --> 00:00:43,467
two question-answering groups, a code-text group, and two agent
groups.
两个问答组、一个代码文本组，以及两个智能体组。

12
00:00:44,287 --> 00:00:48,687
The grouping follows the mean content embeddings of
representative records from each source,
分组依据各来源代表记录的平均内容嵌入，

13
00:00:49,167 --> 00:00:51,607
so related sources share a capability theme.
使相关来源共享一个能力主题。

14
00:00:52,427 --> 00:00:57,707
This is grouping sources, not forcing every training record into
one universal quality scale.
这是对来源分组，而不是把每条训练记录强行放进统一质量尺度。

15
00:00:58,547 --> 00:01:02,607
A group gives the next stage a coherent place to discover
evaluation criteria.
一个来源组为下一阶段发现评价准则提供了相对一致的环境。

16
00:01:03,527 --> 00:01:07,607
The paper then uses a separate rubric and a separate distilled
scorer for each group.
论文随后为每组使用独立评分规则和独立蒸馏评分器。

17
00:01:08,233 --> 00:01:11,973
Within a group, a teacher first judges sampled records freely.
在每个来源组内，教师先自由评判抽样记录。

18
00:01:12,593 --> 00:01:19,133
It proposes quality dimensions, assigns scores, and explains its
reasons without receiving a pre-written rubric.
在没有预先编写的评分规则时，它提出质量维度、给出分数，并解释原因。

19
00:01:20,333 --> 00:01:25,713
Each dimension name and reason becomes a judgment point that can
be embedded and clustered.
每个维度名称及其原因构成一个判断点，可以进行嵌入和聚类。

20
00:01:26,753 --> 00:01:30,473
For each cluster, the point nearest its center becomes an anchor
dimension.
每个聚类中，最靠近中心的判断点成为锚定维度。

21
00:01:31,333 --> 00:01:34,413
Together these anchors form a fixed vocabulary for the group.
这些锚点共同形成该组固定的评价词汇。

22
00:01:35,313 --> 00:01:43,013
In our illustration, labels such as logical consistency or
observation-sensitive action are examples of questions;
示意中的“逻辑一致性”或“根据观察行动”等标签，只是问题示例；

23
00:01:43,553 --> 00:01:47,513
they are not claimed to be the exact discovered rubric from a
particular source.
并非某来源实际发现的完整评分规则。

24
00:01:48,113 --> 00:01:54,053
Free-form dimensions are useful for discovery, but inconsistent
dimensions cannot directly train a predictable scorer.
自由形式的维度适合发现规则，但不一致的维度不能直接训练稳定评分器。

25
00:01:54,973 --> 00:01:59,513
MIRA freezes the anchors and asks the teacher to score a larger
sample on every anchor,
MIRA 固定锚点，让教师按每个锚点对更大样本评分，

26
00:01:59,933 --> 00:02:03,453
producing a numerical value and a short rationale for each
dimension.
为每个维度生成数值与简短说明。

27
00:02:04,273 --> 00:02:08,713
The anchored labels are split into student-training data and
held-out reliability data.
锚定标签分成学生训练数据和留出的可靠性验证数据。

28
00:02:09,393 --> 00:02:12,893
A separate student for each group learns to emit those scores
and rationales.
每组独立学生模型学习输出这些分数与说明。

29
00:02:13,573 --> 00:02:18,693
The student can then process the full corpus without asking the
expensive teacher to judge every record.
随后学生可以处理完整语料，不需要昂贵教师逐条评判所有记录。

30
00:02:19,293 --> 00:02:23,953
One student may imitate the teacher well on one source and
poorly on another.
同一个学生模型可能在一个来源上很好地模仿教师，在另一个来源上却较差。

31
00:02:25,033 --> 00:02:30,353
MIRA checks that agreement on held-out records, separately for
each source and anchor dimension.
MIRA 在留出记录上检查这种一致性，并分别考察每个来源和锚定维度。

32
00:02:31,393 --> 00:02:37,713
Mean absolute error and rank correlation identify
source-dimension pairs whose scores are unreliable.
平均绝对误差和秩相关用于识别分数不可靠的来源与维度组合。

33
00:02:38,853 --> 00:02:40,593
The correction happens after scoring.
修正发生在评分之后。

34
00:02:41,833 --> 00:02:47,733
An unreliable dimension is masked for that source; a trimmed
mean combines the surviving dimensions.
对于某个来源，不可靠维度被屏蔽；剩余维度通过截尾均值合并。

35
00:02:49,073 --> 00:02:55,773
The student prompt stays fixed, and full-corpus inference need
not be repeated just to update this reliability mask.
学生提示保持不变，更新可靠性掩码也不需要重新对完整语料推理。

36
00:02:56,467 --> 00:02:59,247
Scores also have different distributions across sources.
不同来源的分数分布也不同。

37
00:02:59,887 --> 00:03:04,147
A single global cutoff can remove lower-scoring capability
regions before the others.
单一全局阈值可能先移除分数较低的能力区域。

38
00:03:05,007 --> 00:03:07,867
MIRA therefore compares three retention granularities:
因此，MIRA 比较三种保留粒度：

39
00:03:08,567 --> 00:03:14,347
one global threshold, thresholds inside groups, and thresholds
inside each individual source.
一个全局阈值、各组内部阈值，以及各来源内部阈值。

40
00:03:15,067 --> 00:03:18,067
The best aggregate result reported here is the group variant.
这里报告的最佳综合结果来自组级变体。

41
00:03:18,607 --> 00:03:22,907
It preserves broad capability coverage while allowing sources
within a group to compete.
它保留广泛能力覆盖，同时允许同一组内的来源竞争。

42
00:03:23,867 --> 00:03:28,287
Source-conditioned reliability and source-specific final cutoffs
are related ideas,
按来源检查可靠性与每来源最终阈值，是相关思想，

43
00:03:28,647 --> 00:03:30,367
but they are not the same step of the pipeline.
但不是流程中的同一步。

44
00:03:30,973 --> 00:03:35,813
The experiment uses Qwen two point five Coder with fourteen
billion parameters.
实验使用一百四十亿参数的 Qwen2.5-Coder。

45
00:03:36,713 --> 00:03:42,773
Every checkpoint receives the same subsequent supervised
fine-tuning on four hundred thousand instruction examples.
每个检查点随后都使用同样的四十万条指令样本进行监督微调。

46
00:03:43,873 --> 00:03:46,493
Selection methods train on twenty-five billion tokens,
各选择方法使用二百五十亿个 token 训练，

47
00:03:46,893 --> 00:03:50,333
compared with fifty billion for the unfiltered raw-mixture run.
相比之下，未过滤的原始混合语料运行用五百亿个 token。

48
00:03:51,533 --> 00:03:53,673
Evaluation covers nine code benchmarks.
评估覆盖九个代码基准。

49
00:03:54,533 --> 00:04:02,093
They feed four categories: code generation, multilingual code,
executable SQL, and software engineering.
它们归入四类：代码生成、多语言代码、可执行 SQL 与软件工程。

50
00:04:02,953 --> 00:04:06,433
The final macro average weights those four categories equally;
最终宏平均对四类赋予相同权重；

51
00:04:06,853 --> 00:04:11,393
it is not a pooled success rate or a simple average of all nine
individual scores.
不是合并任务成功率，也不是九个单项分数的简单平均。

52
00:04:12,113 --> 00:04:17,213
On that aggregate, the fifty-billion-token raw mixture scores
sixty-three point eight three.
在该综合指标上，五百亿 token 的原始混合语料得分为六十三点八三。

53
00:04:18,253 --> 00:04:23,793
Source-preserving random selection at twenty-five billion tokens
scores sixty-three point two three.
保留来源分布的随机选择，在二百五十亿 token 下得分六十三点二三。

54
00:04:24,813 --> 00:04:30,113
MIRA Group at the same twenty-five-billion-token budget scores
sixty-four point two zero.
MIRA-Group 在相同的二百五十亿 token 预算下得分六十四点二零。

55
00:04:30,933 --> 00:04:36,993
Using half the training tokens, it reports an aggregate zero
point three seven points above the full-budget run.
使用一半训练 token，报告综合得分比完整预算运行高零点三七分。

56
00:04:37,813 --> 00:04:40,193
The result does not say every benchmark improves.
结果并不表示每个基准都提升。

57
00:04:40,993 --> 00:04:46,573
Its useful evidence is that source-adaptive semantic selection
can preserve the overall capability balance
有用证据是按来源适配的语义选择可保留整体能力平衡，

58
00:04:46,633 --> 00:04:49,053
in this matched code-oriented evaluation.
对应的是这一匹配代码评估设置。

59
00:04:49,620 --> 00:04:51,640
Return to the three original records.
回到三个原创记录。

60
00:04:52,040 --> 00:04:56,140
A derivation, a document, and an agent trace need different
questions.
推导、文档和智能体轨迹需要不同的问题。

61
00:04:57,040 --> 00:05:01,300
The point is not to invent a beautiful universal rubric and hope
it works everywhere.
重点不是设计漂亮通用规则，再希望它适用于所有内容。

62
00:05:02,080 --> 00:05:07,200
First discover a stable quality vocabulary for records with a
shared capability role.
应先为能力角色相近的记录发现稳定质量词汇。

63
00:05:08,240 --> 00:05:11,860
MIRA separates that semantic work from the scalable machinery:
MIRA 把语义工作与可扩展机制分开：

64
00:05:12,660 --> 00:05:17,700
anchor the teacher's judgments, distill a student, diagnose
source-level reliability,
锚定教师判断、蒸馏学生、诊断来源级可靠性，

65
00:05:18,280 --> 00:05:21,020
and choose data at an appropriate retention granularity.
并按合适的保留粒度选择数据。

66
00:05:21,880 --> 00:05:27,780
The training mixture is not just a pile of high scores; it is a
balance of capabilities the next model must learn.
训练混合语料不仅是一堆高分记录，也是下一模型需要学习的能力平衡。