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-both-20260730-v8/dual_round1_review.json

@@ -0,0 +1,30 @@
+{
+  "overall_score": 80,
+  "pass": false,
+  "issues": [
+    {
+      "page": 3,
+      "issue_type": "readability",
+      "severity": "low",
+      "suggestion": "将中英文混合的\"辅助\"统一为英文\"aids\"或中文\"辅助\"，保持语言一致性"
+    },
+    {
+      "page": 13,
+      "issue_type": "image_size",
+      "severity": "medium",
+      "suggestion": "放大Long-Distance Attention图示，确保文本标签清晰可辨，避免因图片过小影响观众理解注意力模式"
+    },
+    {
+      "page": 14,
+      "issue_type": "image_size",
+      "severity": "medium",
+      "suggestion": "增大Anaphora Attention图示尺寸，确保句子中的单词和连接线清晰可见，提升注意力可视化的可读性"
+    },
+    {
+      "page": 17,
+      "issue_type": "overcrowded",
+      "severity": "medium",
+      "suggestion": "将Future directions拆分为两条独立要点：\"Extend to multi-modal inputs; develop local restricted attention\"和\"Explore non-sequential generation; enhance attention interpretability\"，避免单条要点包含过多信息"
+    }
+  ]
+}

chapter5/paper-to-ppt/validation/runs/exp5-4-real-pdf-both-20260730-v8/dual_round1_slides.md

@@ -0,0 +1,138 @@
+---
+theme: default
+---
+
+# Attention Is All You Need
+## Ashish Vaswani et al.
+### NIPS 2017
+
+---
+
+## Abstract
+- First sequence transduction model based solely on attention
+- Dispenses with recurrence and convolutions entirely
+- Superior quality with better parallelization and lower training cost
+- State-of-the-art results: 28.4 BLEU (EN-DE) and 41.8 BLEU (EN-FR)
+
+---
+
+## Background & Key Innovation
+- **Challenges**: RNNs have sequential computation; CNNs need O(log n) layers for long dependencies
+- **Previous work**: Attention typically辅助 recurrent networks
+- **Transformer**: Replaces recurrence/conv with self-attention for global dependencies
+- **Advantages**: Massive parallelization, constant operations for long-range dependencies
+
+---
+
+## Model Architecture Overview
+- Encoder-decoder structure with 6 stacked layers each
+- Encoder: Multi-Head Self-Attention + Feed-Forward Network
+- Decoder: Masked Self-Attention + Encoder-Decoder Attention + Feed-Forward
+- Residual connections and layer normalization; d_model = 512
+
+---
+
+# Transformer Architecture (Figure 1)
+<img src="/paper_figure_1_transformer.png" class="max-h-[500px] w-full object-contain" />
+Complete Transformer architecture with encoder (left) and decoder (right) stacks.
+
+---
+
+## Attention Mechanism & Multi-Head
+- **Scaled Dot-Product Attention**: Attention(Q,K,V) = softmax((QKᵀ)/√dₖ)V
+- **Scaling**: Prevents softmax saturation for large dₖ (dₖ=64)
+- **Multi-Head**: 8 parallel attention heads on projected subspaces
+- **Concatenation**: Combines heads and applies final linear projection
+
+---
+
+## Attention Applications
+1. **Encoder-decoder attention**: Decoder queries attend to encoder outputs
+2. **Encoder self-attention**: All positions attend to each other (global context)
+3. **Decoder self-attention**: Masked to prevent future position access (auto-regressive)
+
+---
+
+## Positional Encoding
+- Injects sequence order information (no recurrence/convolution)
+- Added to input embeddings (same d_model dimension)
+- Uses sine/cosine functions: PE(pos,2i)=sin(pos/10000^(2i/d_model))
+- Alternative: learned embeddings (nearly identical performance)
+
+---
+
+## Self-Attention Advantages
+
+| Aspect               | Self-Attention | RNN        | CNN          |
+|----------------------|----------------|------------|--------------|
+| Complexity           | O(n²·d)        | O(n·d²)    | O(k·n·d²)    |
+| Parallelization      | O(1)           | O(n)       | O(1)         |
+| Long-range paths     | O(1)           | O(n)       | O(logₖn)     |
+
+- Better parallelization than RNNs; shorter paths than CNNs/RNNs
+
+---
+
+## Training Setup
+- Datasets: WMT 2014 EN-DE (4.5M) and EN-FR (36M sentences)
+- Vocabulary: Byte-pair encoding (37K for EN-DE, 32K for EN-FR)
+- Hardware: 8 P100 GPUs; 100K-300K steps (12h-3.5 days)
+- Optimizer: Adam (β₁=0.9, β₂=0.98); learning rate with warmup
+
+---
+
+## Machine Translation Results
+- EN-DE: 28.4 BLEU (↑2+ over previous best, including ensembles)
+- EN-FR: 41.8 BLEU (new single-model state-of-the-art)
+- Training cost: 3.3×10¹⁸ FLOPs (EN-DE base) vs 10¹⁹-10²¹ for competitors
+- Achieves better quality with significantly lower computational resources
+
+---
+
+## Model Variations (Ablation Study)
+- Single-head attention: 0.9 BLEU worse than 8-head setup
+- Reducing dₖ (attention key size) degrades performance
+- Larger models (d_model=1024): +1.1 BLEU over base model
+- Dropout critical for preventing overfitting (P_drop=0.1)
+
+---
+
+# Long-Distance Attention (Figure 3)
+<img src="/paper_figure_3_long_distance.png" class="max-h-[500px] w-full object-contain" />
+Encoder self-attention showing "making" attending to distant "more difficult".
+
+---
+
+# Anaphora Attention (Figure 4)
+<img src="/paper_figure_4_anaphora.png" class="max-h-[500px] w-full object-contain" />
+Attention heads resolving anaphora: "its" attending to "Law" and "application".
+
+---
+
+## Generalization to Parsing
+- Applied Transformer to English constituency parsing
+- WSJ only (40K sentences): 91.3 F1 (comparable to state-of-the-art)
+- Semi-supervised (17M sentences): 92.7 F1 (outperforms most prior models)
+- Demonstrates transferability to structural NLP tasks
+
+---
+
+## Limitations
+- O(n²) self-attention complexity for long sequences
+- Memory-intensive with extended input lengths
+- Decoder still generates output sequentially
+- Attention patterns show interpretability but not fully understood
+
+---
+
+## Conclusion & Future Work
+- **Conclusion**: Establishes new state-of-the-art in translation; replaces recurrence with attention
+- **Future directions**: Extend to multi-modal inputs; develop local restricted attention
+- Explore non-sequential generation; enhance attention interpretability
+- Significantly reduces training time while improving quality
+
+---
+
+## References
+- Vaswani et al. (2017). Attention is all you need. NIPS 2017
+- Code: https://github.com/tensorflow/tensor2tensor

chapter5/paper-to-ppt/validation/runs/exp5-4-real-pdf-both-20260730-v8/dual_round2_review.json

@@ -0,0 +1,18 @@
+{
+  "overall_score": 85,
+  "pass": false,
+  "issues": [
+    {
+      "page": 13,
+      "issue_type": "readability",
+      "severity": "medium",
+      "suggestion": "放大图中文字尺寸，确保观众能清晰辨认注意力可视化图中的词语和连接关系"
+    },
+    {
+      "page": 14,
+      "issue_type": "readability",
+      "severity": "medium",
+      "suggestion": "增大指代注意力图中的文本字号，或局部放大关键区域以提升可读性"
+    }
+  ]
+}

chapter5/paper-to-ppt/validation/runs/exp5-4-real-pdf-both-20260730-v8/dual_round2_slides.md

@@ -0,0 +1,146 @@
+---
+theme: default
+---
+
+# Attention Is All You Need
+## Ashish Vaswani et al.
+### NIPS 2017
+
+---
+
+## Abstract
+- First sequence transduction model based solely on attention
+- Dispenses with recurrence and convolutions entirely
+- Superior quality with better parallelization and lower training cost
+- State-of-the-art results: 28.4 BLEU (EN-DE) and 41.8 BLEU (EN-FR)
+
+---
+
+## Background & Key Innovation
+- **Challenges**: RNNs have sequential computation; CNNs need O(log n) layers for long dependencies
+- **Previous work**: Attention typically aids recurrent networks
+- **Transformer**: Replaces recurrence/conv with self-attention for global dependencies
+- **Advantages**: Massive parallelization, constant operations for long-range dependencies
+
+---
+
+## Model Architecture Overview
+- Encoder-decoder structure with 6 stacked layers each
+- Encoder: Multi-Head Self-Attention + Feed-Forward Network
+- Decoder: Masked Self-Attention + Encoder-Decoder Attention + Feed-Forward
+- Residual connections and layer normalization; d_model = 512
+
+---
+
+# Transformer Architecture (Figure 1)
+<img src="/paper_figure_1_transformer.png" class="max-h-[500px] w-full object-contain" />
+Complete Transformer architecture with encoder (left) and decoder (right) stacks.
+
+---
+
+## Attention Mechanism & Multi-Head
+- **Scaled Dot-Product Attention**: Attention(Q,K,V) = softmax((QKᵀ)/√dₖ)V
+- **Scaling**: Prevents softmax saturation for large dₖ (dₖ=64)
+- **Multi-Head**: 8 parallel attention heads on projected subspaces
+- **Concatenation**: Combines heads and applies final linear projection
+
+---
+
+## Attention Applications
+1. **Encoder-decoder attention**: Decoder queries attend to encoder outputs
+2. **Encoder self-attention**: All positions attend to each other (global context)
+3. **Decoder self-attention**: Masked to prevent future position access (auto-regressive)
+
+---
+
+## Positional Encoding
+- Injects sequence order information (no recurrence/convolution)
+- Added to input embeddings (same d_model dimension)
+- Uses sine/cosine functions: PE(pos,2i)=sin(pos/10000^(2i/d_model))
+- Alternative: learned embeddings (nearly identical performance)
+
+---
+
+## Self-Attention Advantages
+
+| Aspect               | Self-Attention | RNN        | CNN          |
+|----------------------|----------------|------------|--------------|
+| Complexity           | O(n²·d)        | O(n·d²)    | O(k·n·d²)    |
+| Parallelization      | O(1)           | O(n)       | O(1)         |
+| Long-range paths     | O(1)           | O(n)       | O(logₖn)     |
+
+- Better parallelization than RNNs; shorter paths than CNNs/RNNs
+
+---
+
+## Training Setup
+- Datasets: WMT 2014 EN-DE (4.5M) and EN-FR (36M sentences)
+- Vocabulary: Byte-pair encoding (37K for EN-DE, 32K for EN-FR)
+- Hardware: 8 P100 GPUs; 100K-300K steps (12h-3.5 days)
+- Optimizer: Adam (β₁=0.9, β₂=0.98); learning rate with warmup
+
+---
+
+## Machine Translation Results
+- EN-DE: 28.4 BLEU (↑2+ over previous best, including ensembles)
+- EN-FR: 41.8 BLEU (new single-model state-of-the-art)
+- Training cost: 3.3×10¹⁸ FLOPs (EN-DE base) vs 10¹⁹-10²¹ for competitors
+- Achieves better quality with significantly lower computational resources
+
+---
+
+## Model Variations (Ablation Study)
+- Single-head attention: 0.9 BLEU worse than 8-head setup
+- Reducing dₖ (attention key size) degrades performance
+- Larger models (d_model=1024): +1.1 BLEU over base model
+- Dropout critical for preventing overfitting (P_drop=0.1)
+
+---
+
+# Long-Distance Attention (Figure 3)
+<img src="/paper_figure_3_long_distance.png" class="max-h-[500px] w-full object-contain" />
+Encoder self-attention showing "making" attending to distant "more difficult".
+
+---
+
+# Anaphora Attention (Figure 4)
+<img src="/paper_figure_4_anaphora.png" class="max-h-[500px] w-full object-contain" />
+Attention heads resolving anaphora: "its" attending to "Law" and "application".
+
+---
+
+## Generalization to Parsing
+- Applied Transformer to English constituency parsing
+- WSJ only (40K sentences): 91.3 F1 (comparable to state-of-the-art)
+- Semi-supervised (17M sentences): 92.7 F1 (outperforms most prior models)
+- Demonstrates transferability to structural NLP tasks
+
+---
+
+## Limitations
+- O(n²) self-attention complexity for long sequences
+- Memory-intensive with extended input lengths
+- Decoder still generates output sequentially
+- Attention patterns show interpretability but not fully understood
+
+---
+
+## Conclusion
+- Establishes new state-of-the-art in machine translation
+- Replaces recurrence/convolution with attention for better parallelization
+- Generalizes effectively to other sequence tasks (e.g., parsing)
+- Significantly reduces training time while improving quality
+
+---
+
+## Future Work
+- Extend to multi-modal inputs; develop local restricted attention
+- Explore non-sequential generation; enhance attention interpretability
+- Optimize for longer sequences with improved memory efficiency
+- Further investigate attention mechanism interpretability
+
+---
+
+## References
+- Vaswani et al. (2017). Attention is all you need. NIPS 2017
+- Code: https://github.com/tensorflow/tensor2tensor

chapter5/paper-to-ppt/validation/runs/exp5-4-real-pdf-both-20260730-v8/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-both-20260730-v8/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,
+        60,
+        520,
+        405
+      ],
+      "sha256": "09364ee993caf62234733a5aaacc31bc0472b8db4b6107455f898e3a7a020597",
+      "bytes": 94487,
+      "public_copy_sha256": "09364ee993caf62234733a5aaacc31bc0472b8db4b6107455f898e3a7a020597"
+    },
+    {
+      "filename": "paper_figure_3_long_distance.png",
+      "pdf_page": 13,
+      "source_label": "Figure 3 (long-distance dependency focus)",
+      "caption": "Published encoder attention linking 'making' to 'more difficult'.",
+      "rect": [
+        190,
+        88,
+        425,
+        311
+      ],
+      "sha256": "611a465fc6ba0f266c0a4743600b26ba861ebee93597e04f7c34067a8ddba296",
+      "bytes": 43326,
+      "public_copy_sha256": "611a465fc6ba0f266c0a4743600b26ba861ebee93597e04f7c34067a8ddba296"
+    },
+    {
+      "filename": "paper_figure_4_anaphora.png",
+      "pdf_page": 14,
+      "source_label": "Figure 4 (lower panel, anaphora focus)",
+      "caption": "Published attention from 'its' to 'Law' and 'application'.",
+      "rect": [
+        92,
+        360,
+        310,
+        610
+      ],
+      "sha256": "fdc5b28e860e65ff5922d237d190fa11f5586bfb00894b58d09cf38299bfef67",
+      "bytes": 30161,
+      "public_copy_sha256": "fdc5b28e860e65ff5922d237d190fa11f5586bfb00894b58d09cf38299bfef67"
+    }
+  ]
+}