` 包裹；长文使用 `
`/`
` 划分小节。 3. 格式要求：列表用 `
- `，引用用 `
`，代码用 `
`，重要句子用 ``。
4. 内容约束：保留原文所有信息，不得增删改内容。
5. 文档级标签：不要输出 ``、``、`` 等标签。
* 输入文本分析：
* 这是一段关于EEG基础模型评估的技术/科学摘要。
* 它描述了问题（缺乏三种对照）、方法（评估5个公开模型和1个内部模型）、结果（三种变化方式）和结论（提出四项报告实践，发布所有数据）。
2. 将文本拆解为HTML组件：
* *第一句：* "EEG foundation models report downstream gains, yet their protocols seldom include three controls." -> 这是一个关键的背景介绍。可以使用 `EEG foundation models report downstream gains, yet their protocols seldom include three controls.
`。
* *第二句：* "These are the same encoder untrained, spectral features fitted on the same data, and a layer picked without looking at test scores." -> 这是对这三项对照的列举。可以使用 `These are
- the same encoder untrained
- spectral features fitted on the same data
- a layer picked without looking at test scores
` 或者将其拆分。等等，HTML语义化中 `` 不能直接嵌套在 `` 内部（会导致解析错误）。最好用 `
- ...
` 独立放置，或者通过引入 `` 来组织。让我们回顾一下自然段落。
* *第三句：* "We therefore evaluate five publicly released EEG foundation models and an in-house JEPA-style model under one protocol with all three controls and several seeds, on five benchmarks and four added task families." -> 这是方法论/实验设置。可以使用 `
Evaluation Protocol
We therefore evaluate...
`。
* *第四句：* "The controls change the reading in three ways." -> 引入结果部分。可以使用 `Results
The controls change the reading in three ways.
`。
* *第五至第七句：* 描述了三种变化（First, Second, Third）。
* First: "First, picking the probed layer by its test score raises the layer-wise peak by an average of +0.034 and reverses the peak-versus-output gap in 9 of 36 cells." -> `- First: picking the probed layer...
`
* Second: "Second, with the layer picked on validation, random initialisation matches or beats pretraining for half of the models on sleep and emotion, and no model beats a five-band log-power baseline on emotion or motor imagery." -> `- Second: with the layer...
`
* Third: "Third, after full fine-tuning only 5 of 30 pairs gain beyond twice their seed spread, and the one clear loss of -0.290 AUROC persists under matched learning rates." -> `- Third: after full fine-tuning...
`
* *最后一句：* "From these results we draw four reporting practices, and we release every cell for re-analysis." -> 结论/结论。可以使用 `Conclusion
From these results...
`。
3. 优化HTML结构（同时严格遵守“不增删内容”的规则）：
* 等等，增加 `Evaluation Protocol
` 算不算“增加内容”？规则说明“保留原文所有信息，不得增删改内容”。增加标题可能会在严格意义上被视为增加了并非原文的词汇。如果可能的话，为了绝对安全，我应该避免增加标题，或者只在原文确实像长篇报告的特定章节那样应用它们。既然这是一段较短的摘要（一段很长的文本），我最好直接通过 `` 和列表来组织结构，不增加额外的标题文本，以避免违反“不增删内容”的规则。
* 等等，如果增加 `
` 或 `
` 标签，没有可见文本算作“增加内容”吗？增加 `
Evaluation Protocol
` 会增加可见文本。如果我只增加 ``，那在语义上毫无意义。让我们重新阅读规则：“保留原文所有信息，不得增删改内容”。这意味着不要增加句子。我如果全部使用 `` 和 `
` 标签，不增加任何标题，可能更安全，或者只在为原已存在的列表建立分组时才使用标题。
* 让我们看看列表部分：“These are the same encoder untrained, spectral features fitted on the same data, and a layer picked without looking at test scores.” -> 转换为 `` 会不会改变措辞？我需要保留原话。如果我把它转换为列表，我必须把它拆分开。
* "These are" -> `These are
`
- the same encoder untrained
- spectral features fitted on the same data
- a layer picked without looking at test scores
* “First, Second, Third” 部分呢？
* 原文：“First, picking the probed layer by its test score raises the layer-wise peak by an average of +0.034 and reverses the peak-versus-output gap in 9 of 36 cells. Second, with the layer picked on validation, random initialisation matches or beats pretraining for half of the models on sleep and emotion, and no model beats a five-band log-power baseline on emotion or motor imagery. Third, after full fine-tuning only 5 of 30 pairs gain beyond twice their seed spread, and the one clear loss of -0.290 AUROC persists under matched learning rates.”
* 我可以将其放入列表中：
- First, picking the probed layer by its test score raises the layer-wise peak by an average of +0.034 and reverses the peak-versus-output gap in 9 of 36 cells.
- Second, with the layer picked on validation, random initialisation matches or beats pretraining for half of the models on sleep and emotion, and no model beats a five-band log-power baseline on emotion or motor imagery.
- Third, after full fine-tuning only 5 of 30 pairs gain beyond twice their seed spread, and the one clear loss of -0.290 AUROC persists under matched learning rates.
* 让我们重新检查整个文本以进行逐字拆分。
原文：
"EEG foundation models report downstream gains, yet their protocols seldom include three controls. These are the same encoder untrained, spectral features fitted on the same data, and a layer picked without looking at test scores. We therefore evaluate five publicly released EEG foundation models and an in-house JEPA-style model under one protocol with all three controls and several seeds, on five benchmarks and four added task families. The controls change the reading in three ways. First, picking the probed layer by its test score raises the layer-wise peak by an average of +0.034 and reverses the peak-versus-output gap in 9 of 36 cells. Second, with the layer picked on validation, random initialisation matches or beats pretraining for half of the models on sleep and emotion, and no model beats a five-band log-power baseline on emotion or motor imagery. Third, after full fine-tuning only 5 of 30 pairs gain beyond twice their seed spread, and the one clear loss of -0.290 AUROC persists under matched learning rates. From these results we draw four reporting practices, and we release every cell for re-analysis."
让我们进行切分：
第1段（背景）：
EEG foundation models report downstream gains, yet their
`，代码用 ``，重要句子用 ``。 4. 内容约束：保留原文所有信息，不得增删改内容。 5. 文档级标签：不要输出 ``、``、`` 等标签。 * 输入文本分析： * 这是一段关于EEG基础模型评估的技术/科学摘要。 * 它描述了问题（缺乏三种对照）、方法（评估5个公开模型和1个内部模型）、结果（三种变化方式）和结论（提出四项报告实践，发布所有数据）。 2. 将文本拆解为HTML组件： * *第一句：* "EEG foundation models report downstream gains, yet their protocols seldom include three controls." -> 这是一个关键的背景介绍。可以使用 `EEG foundation models report downstream gains, yet their protocols seldom include three controls.
`。 * *第二句：* "These are the same encoder untrained, spectral features fitted on the same data, and a layer picked without looking at test scores." -> 这是对这三项对照的列举。可以使用 `These are
` 或者将其拆分。等等，HTML语义化中 `
- the same encoder untrained
- spectral features fitted on the same data
- a layer picked without looking at test scores
` 不能直接嵌套在 `
` 内部（会导致解析错误）。最好用 `
` 独立放置，或者通过引入 `
- ...
` 来组织。让我们回顾一下自然段落。 * *第三句：* "We therefore evaluate five publicly released EEG foundation models and an in-house JEPA-style model under one protocol with all three controls and several seeds, on five benchmarks and four added task families." -> 这是方法论/实验设置。可以使用 `
Evaluation Protocol
We therefore evaluate...
`。 * *第四句：* "The controls change the reading in three ways." -> 引入结果部分。可以使用 `Results
The controls change the reading in three ways.
`。 * *第五至第七句：* 描述了三种变化（First, Second, Third）。 * First: "First, picking the probed layer by its test score raises the layer-wise peak by an average of +0.034 and reverses the peak-versus-output gap in 9 of 36 cells." -> `` * Second: "Second, with the layer picked on validation, random initialisation matches or beats pretraining for half of the models on sleep and emotion, and no model beats a five-band log-power baseline on emotion or motor imagery." -> `
- First: picking the probed layer...
- Second: with the layer...
` * Third: "Third, after full fine-tuning only 5 of 30 pairs gain beyond twice their seed spread, and the one clear loss of -0.290 AUROC persists under matched learning rates." -> `- Third: after full fine-tuning...
` * *最后一句：* "From these results we draw four reporting practices, and we release every cell for re-analysis." -> 结论/结论。可以使用 `Conclusion
From these results...
`。 3. 优化HTML结构（同时严格遵守“不增删内容”的规则）： * 等等，增加 `Evaluation Protocol
` 算不算“增加内容”？规则说明“保留原文所有信息，不得增删改内容”。增加标题可能会在严格意义上被视为增加了并非原文的词汇。如果可能的话，为了绝对安全，我应该避免增加标题，或者只在原文确实像长篇报告的特定章节那样应用它们。既然这是一段较短的摘要（一段很长的文本），我最好直接通过 `` 和列表来组织结构，不增加额外的标题文本，以避免违反“不增删内容”的规则。 * 等等，如果增加 `
` 或 `
` 标签，没有可见文本算作“增加内容”吗？增加 `
Evaluation Protocol
` 会增加可见文本。如果我只增加 ``，那在语义上毫无意义。让我们重新阅读规则：“保留原文所有信息，不得增删改内容”。这意味着不要增加句子。我如果全部使用 `` 和 `
` 标签，不增加任何标题，可能更安全，或者只在为原已存在的列表建立分组时才使用标题。 * 让我们看看列表部分：“These are the same encoder untrained, spectral features fitted on the same data, and a layer picked without looking at test scores.” -> 转换为 `
` 会不会改变措辞？我需要保留原话。如果我把它转换为列表，我必须把它拆分开。 * "These are" -> `
These are
`* “First, Second, Third” 部分呢？ * 原文：“First, picking the probed layer by its test score raises the layer-wise peak by an average of +0.034 and reverses the peak-versus-output gap in 9 of 36 cells. Second, with the layer picked on validation, random initialisation matches or beats pretraining for half of the models on sleep and emotion, and no model beats a five-band log-power baseline on emotion or motor imagery. Third, after full fine-tuning only 5 of 30 pairs gain beyond twice their seed spread, and the one clear loss of -0.290 AUROC persists under matched learning rates.” * 我可以将其放入列表中：
- the same encoder untrained
- spectral features fitted on the same data
- a layer picked without looking at test scores
* 让我们重新检查整个文本以进行逐字拆分。 原文： "EEG foundation models report downstream gains, yet their protocols seldom include three controls. These are the same encoder untrained, spectral features fitted on the same data, and a layer picked without looking at test scores. We therefore evaluate five publicly released EEG foundation models and an in-house JEPA-style model under one protocol with all three controls and several seeds, on five benchmarks and four added task families. The controls change the reading in three ways. First, picking the probed layer by its test score raises the layer-wise peak by an average of +0.034 and reverses the peak-versus-output gap in 9 of 36 cells. Second, with the layer picked on validation, random initialisation matches or beats pretraining for half of the models on sleep and emotion, and no model beats a five-band log-power baseline on emotion or motor imagery. Third, after full fine-tuning only 5 of 30 pairs gain beyond twice their seed spread, and the one clear loss of -0.290 AUROC persists under matched learning rates. From these results we draw four reporting practices, and we release every cell for re-analysis." 让我们进行切分： 第1段（背景）： EEG foundation models report downstream gains, yet their
- First, picking the probed layer by its test score raises the layer-wise peak by an average of +0.034 and reverses the peak-versus-output gap in 9 of 36 cells.
- Second, with the layer picked on validation, random initialisation matches or beats pretraining for half of the models on sleep and emotion, and no model beats a five-band log-power baseline on emotion or motor imagery.
- Third, after full fine-tuning only 5 of 30 pairs gain beyond twice their seed spread, and the one clear loss of -0.290 AUROC persists under matched learning rates.