ai-agent-book 精选快照（<2MB 代码与文档，来自 github.com/bojieli/ai-agent-book）

chapter5/paper-to-ppt/validation/runs/exp5-4-real-pdf-20260730-v3/dual_round1_review.json

@@ -0,0 +1,66 @@
+{
+  "overall_score": 55,
+  "pass": false,
+  "issues": [
+    {
+      "page": 2,
+      "issue_type": "overcrowded",
+      "severity": "medium",
+      "suggestion": "将6个要点拆分为2页，例如将BLEU分数单独作为一页突出展示关键结果"
+    },
+    {
+      "page": 3,
+      "issue_type": "overcrowded",
+      "severity": "medium",
+      "suggestion": "将传统方法的挑战拆分为单独页面，或减少子要点数量至2个以内"
+    },
+    {
+      "page": 4,
+      "issue_type": "overcrowded",
+      "severity": "medium",
+      "suggestion": "将现有模型局限性和注意力机制分为2页，每部分保留3个核心要点"
+    },
+    {
+      "page": 5,
+      "issue_type": "image_size",
+      "severity": "high",
+      "suggestion": "增大架构图尺寸并移除图片下方的文字说明，或单独用一页展示清晰的架构图"
+    },
+    {
+      "page": 5,
+      "issue_type": "readability",
+      "severity": "high",
+      "suggestion": "确保架构图中的文字清晰可辨，必要时拆分图表或局部放大关键组件"
+    },
+    {
+      "page": 5,
+      "issue_type": "overcrowded",
+      "severity": "medium",
+      "suggestion": "删除图片下方的长段文字说明，仅保留底部的总结句"
+    },
+    {
+      "page": 6,
+      "issue_type": "overcrowded",
+      "severity": "medium",
+      "suggestion": "将编码器架构的子层信息拆分为单独页面，或简化为要点列表"
+    },
+    {
+      "page": 7,
+      "issue_type": "overcrowded",
+      "severity": "medium",
+      "suggestion": "将解码器的三个子层拆分为单独页面详细说明，主页面仅保留核心结构概述"
+    },
+    {
+      "page": 8,
+      "issue_type": "overcrowded",
+      "severity": "low",
+      "suggestion": "缩短定义部分的描述文字，或拆分为单独的'注意力定义'和'注意力类型'两页"
+    },
+    {
+      "page": 10,
+      "issue_type": "layout",
+      "severity": "high",
+      "suggestion": "补充Multi-Head Attention的核心要点（如并行注意力头、拼接输出等），避免空页"
+    }
+  ]
+}

chapter5/paper-to-ppt/validation/runs/exp5-4-real-pdf-20260730-v3/dual_round1_slides.md

@@ -0,0 +1,109 @@
+---
+theme: default
+title: "Attention Is All You Need"
+author: "Ashish Vaswani et al."
+---
+
+# Attention Is All You Need
+
+**Authors:** Ashish Vaswani et al.  
+**Conference:** NIPS 2017
+
+---
+
+## Abstract
+
+- Proposes **Transformer** - a new network architecture based solely on attention
+- Dispenses with recurrence and convolutions entirely
+- More parallelizable and requires significantly less training time
+- Achieves state-of-the-art results on machine translation tasks:
+  - 28.4 BLEU on WMT 2014 English-to-German
+  - 41.8 BLEU on WMT 2014 English-to-French
+- Generalizes well to other tasks like English constituency parsing
+
+---
+
+## Background: Sequence Modeling Challenges
+
+### Traditional Approaches
+- **Recurrent models** (LSTM, GRU) process input sequentially
+- **Convolutional models** have limited receptive fields
+- Both struggle with:
+  - Long-range dependencies
+  - Parallelization
+  - Computational efficiency for long sequences
+
+---
+
+## The Need for a New Architecture
+
+### Key Limitations of Existing Models
+- **Recurrent networks**: Inherently sequential computation
+- **Convolutional networks**: Limited long-range connectivity
+- **Hybrid models**: Still rely on recurrence/convolution as primary components
+
+### Attention Mechanisms
+- Already used as auxiliary component in sequence models
+- Allows modeling dependencies without regard to position distance
+- Not yet used as the primary architectural component
+
+---
+
+## The Transformer: Model Architecture
+
+<img src="/paper_figure_1_transformer.png" class="h-76 mx-auto" />
+
+*Encoder-decoder structure using stacked self-attention and feed-forward layers*
+
+---
+
+## Encoder Architecture
+
+- **Stack of 6 identical layers**
+- Each layer contains two sub-layers:
+  1. Multi-head self-attention mechanism
+  2. Position-wise fully connected feed-forward network
+- Residual connections around each sub-layer
+- Layer normalization: `LayerNorm(x + Sublayer(x))`
+- Output dimension: `d_model = 512`
+
+---
+
+## Decoder Architecture
+
+- **Stack of 6 identical layers**
+- Each layer contains three sub-layers:
+  1. Masked multi-head self-attention
+  2. Multi-head attention over encoder output
+  3. Position-wise fully connected feed-forward network
+- Residual connections and layer normalization
+- Masking prevents attending to future positions
+
+---
+
+## Attention Mechanism Basics
+
+### Definition
+An attention function maps a query and key-value pairs to an output:
+- Query (Q), Keys (K), Values (V) are vectors
+- Output is weighted sum of values
+- Weights determined by compatibility of query with corresponding key
+
+### Two Common Types
+- Additive attention (feed-forward network)
+- Dot-product attention (scaled in Transformer)
+
+---
+
+## Scaled Dot-Product Attention
+
+$$\text{Attention}(Q, K, V) = \text{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right)V$$
+
+- Computes dot products of query with all keys
+- Scales by $\sqrt{d_k}$ to prevent gradient vanishing
+- Applies softmax to get weights over values
+- More efficient than additive attention (matrix operations)
+
+---
+
+## Multi-Head Attention

chapter5/paper-to-ppt/validation/runs/exp5-4-real-pdf-20260730-v3/dual_round2_review.json

@@ -0,0 +1,30 @@
+{
+  "overall_score": 85,
+  "pass": false,
+  "issues": [
+    {
+      "page": 4,
+      "issue_type": "overcrowded",
+      "severity": "medium",
+      "suggestion": "将6个要点拆分为两页，Recurrent Models和Convolutional Models各占一页，或删减次要内容使每页不超过5个要点"
+    },
+    {
+      "page": 6,
+      "issue_type": "overcrowded",
+      "severity": "high",
+      "suggestion": "移除图片下方的大段文字说明，仅保留模型图；文字内容应拆分到后续页面详细讲解"
+    },
+    {
+      "page": 7,
+      "issue_type": "overcrowded",
+      "severity": "low",
+      "suggestion": "将\"Each layer contains two sub-layers\"的两个子要点合并为一条，或调整字号增大留白"
+    },
+    {
+      "page": 9,
+      "issue_type": "overcrowded",
+      "severity": "low",
+      "suggestion": "将三个子要点简化为更简短的表述，或拆分到新页面单独说明Decoder的子层结构"
+    }
+  ]
+}

chapter5/paper-to-ppt/validation/runs/exp5-4-real-pdf-20260730-v3/dual_round2_slides.md

@@ -0,0 +1,126 @@
+---
+theme: default
+title: "Attention Is All You Need"
+author: "Ashish Vaswani et al."
+---
+
+# Attention Is All You Need
+
+**Authors:** Ashish Vaswani et al.  
+**Conference:** NIPS 2017
+
+---
+
+## Abstract: Key Innovation
+
+- Proposes **Transformer** - first architecture based solely on attention
+- Dispenses with recurrence and convolutions entirely
+- More parallelizable and requires significantly less training time
+- Generalizes well to other tasks beyond machine translation
+
+---
+
+## Abstract: Performance Highlights
+
+- **WMT 2014 English-to-German**: 28.4 BLEU
+  - Improves over existing best results by over 2 BLEU
+- **WMT 2014 English-to-French**: 41.8 BLEU
+  - New single-model state-of-the-art
+  - Trained for 3.5 days on eight GPUs (small fraction of previous costs)
+
+---
+
+## Background: Traditional Sequence Modeling
+
+### Recurrent Models
+- LSTM and GRU as state-of-the-art approaches
+- Factor computation along symbol positions
+- Inherently sequential nature precludes parallelization
+
+### Convolutional Models
+- Fixed receptive fields limit long-range dependencies
+- Computational complexity grows with distance between positions
+
+---
+
+## Limitations of Existing Approaches
+
+- **Recurrent networks**: Sequential computation limits parallelization
+- **Convolutional networks**: Difficulty modeling long-range dependencies
+- **Hybrid models**: Still rely on recurrence/convolution as primary components
+- **Attention mechanisms**: Previously used only as auxiliary component
+
+---
+
+## The Transformer: Model Architecture
+
+<img src="/paper_figure_1_transformer.png" class="h-80 mx-auto" />
+
+---
+
+## Encoder Architecture
+
+- **Stack of 6 identical layers**
+- Each layer contains two sub-layers:
+  1. Multi-head self-attention mechanism
+  2. Position-wise fully connected feed-forward network
+- Residual connections around each sub-layer
+- Layer normalization: `LayerNorm(x + Sublayer(x))`
+
+---
+
+## Encoder: Key Details
+
+- All sub-layers produce outputs of dimension `d_model = 512`
+- Self-attention allows each position to attend to all positions
+- Feed-forward network applied to each position separately
+- Residual connections help with gradient flow in deep networks
+
+---
+
+## Decoder Architecture
+
+- **Stack of 6 identical layers**
+- Each layer contains three sub-layers:
+  1. Masked multi-head self-attention
+  2. Multi-head attention over encoder output
+  3. Position-wise fully connected feed-forward network
+- Residual connections and layer normalization
+
+---
+
+## Decoder: Masking Mechanism
+
+- **Masked self-attention** prevents positions from attending to subsequent positions
+- Ensures predictions for position i depend only on:
+  - Known outputs at positions less than i
+  - Input sequence through encoder-decoder attention
+- Output embeddings offset by one position to maintain auto-regressive property
+
+---
+
+## Attention Mechanism: Definition
+
+Maps a query and key-value pairs to an output:
+- Query (Q), Keys (K), Values (V) are vectors
+- Output is weighted sum of values
+- Weights determined by query-key compatibility
+
+---
+
+## Scaled Dot-Product Attention
+
+$$\text{Attention}(Q, K, V) = \text{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right)V$$
+
+- Computes dot products of query with all keys
+- Scales by $\sqrt{d_k}$ to prevent gradient vanishing
+- Applies softmax to get weights over values
+- More efficient than additive attention
+
+---
+
+## Multi-Head Attention
+
+- Projects Q, K, V h times with different learned projections
+- Performs attention in parallel on each projected version
+- Concatenates outputs and applies final projection

chapter5/paper-to-ppt/validation/runs/exp5-4-real-pdf-20260730-v3/source/source_visuals/manifest.json

@@ -0,0 +1,62 @@
+{
+  "paper": {
+    "title": "Attention Is All You Need",
+    "authors": "Ashish Vaswani et al.",
+    "arxiv_id": "1706.03762",
+    "pdf_url": "https://arxiv.org/pdf/1706.03762",
+    "pdf_sha256": "bdfaa68d8984f0dc02beaca527b76f207d99b666d31d1da728ee0728182df697",
+    "observed_pdf_sha256": "bdfaa68d8984f0dc02beaca527b76f207d99b666d31d1da728ee0728182df697"
+  },
+  "paper_text": {
+    "path": "/Users/boj/book/ai-agent-book/chapter5/paper-to-ppt/validation/runs/exp5-4-real-pdf-20260730-v3/source/paper_text.md",
+    "sha256": "4da5f2a1da38bad8149832b5ada3c94f0619188db8fed5cdca8ff63405587eb8",
+    "characters": 39819
+  },
+  "visuals": [
+    {
+      "filename": "paper_figure_1_transformer.png",
+      "pdf_page": 3,
+      "source_label": "Figure 1",
+      "caption": "The Transformer model architecture.",
+      "rect": [
+        92,
+        28,
+        520,
+        535
+      ],
+      "sha256": "48f09e2e2426e8ca76c1e438600438dca2f9e61360e92ae3edbb538a2cc611d2",
+      "bytes": 150087,
+      "public_copy_sha256": "48f09e2e2426e8ca76c1e438600438dca2f9e61360e92ae3edbb538a2cc611d2"
+    },
+    {
+      "filename": "paper_figure_3_long_distance.png",
+      "pdf_page": 13,
+      "source_label": "Figure 3",
+      "caption": "Encoder self-attention following long-distance dependencies.",
+      "rect": [
+        92,
+        55,
+        525,
+        455
+      ],
+      "sha256": "ea456861962564d2ef5a158e4205cfea29b537f4a611df989db5d65471f23d78",
+      "bytes": 96204,
+      "public_copy_sha256": "ea456861962564d2ef5a158e4205cfea29b537f4a611df989db5d65471f23d78"
+    },
+    {
+      "filename": "paper_figure_4_anaphora.png",
+      "pdf_page": 14,
+      "source_label": "Figure 4",
+      "caption": "Attention heads involved in anaphora resolution.",
+      "rect": [
+        92,
+        135,
+        525,
+        665
+      ],
+      "sha256": "b1f0c6b0e186acb50cae69c304a9a2a0e1f09fd15f1beb0bba03194528163d7e",
+      "bytes": 264855,
+      "public_copy_sha256": "b1f0c6b0e186acb50cae69c304a9a2a0e1f09fd15f1beb0bba03194528163d7e"
+    }
+  ]
+}