## Heatmap: Token-Layer Attention Distribution ### Overview The image is a heatmap visualizing the distribution of attention weights or similarity scores between input tokens and transformer model layers. The x-axis represents input tokens (e.g., question/answer components), while the y-axis represents model layers (0–30). Darker blue shades indicate higher values (closer to 1.0), and lighter shades represent lower values (closer to 0.5). --- ### Components/Axes - **X-axis (Tokens)**: - `last_q` - `first_answer` - `second_answer` - `exact_answer_before_first` - `exact_answer_first` - `exact_answer_last` - `exact_answer_after_last` - **Y-axis (Layers)**: - Layer indices: 0 (bottom) to 30 (top) - **Color Scale**: - Legend on the right: Dark blue (1.0) to light gray (0.5) - **Spatial Layout**: - Legend positioned vertically on the right side of the heatmap. - Tokens labeled at the bottom, layers labeled on the left. --- ### Detailed Analysis 1. **Token-Layer Patterns**: - **`last_q`**: High values (dark blue) concentrated in layers 28–30, suggesting strong attention to the final question token in later layers. - **`first_answer`**: Peaks in layers 12–16, with moderate values in layers 18–22. - **`second_answer`**: Similar to `first_answer`, with peaks in layers 12–16 and 18–22. - **`exact_answer_before_first`**: High values in layers 24–28, indicating late-layer focus. - **`exact_answer_first`**: Peaks in layers 24–28, with gradual decline toward layer 30. - **`exact_answer_last`**: Strongest values in layers 28–30, mirroring `last_q`. - **`exact_answer_after_last`**: High values in layers 28–30, similar to `exact_answer_last`. 2. **Value Distribution**: - Most tokens show elevated values in mid-to-late layers (12–30), with the highest concentrations in layers 24–30. - Early layers (0–11) exhibit uniformly low values (<0.6) across all tokens. --- ### Key Observations - **Layer-Specific Attention**: - Early layers (0–11) show minimal engagement with all tokens, suggesting initial processing focuses on basic tokenization or positional encoding. - Mid-layers (12–22) handle answer-related tokens (`first_answer`, `second_answer`), while late layers (24–30) dominate for question and exact answer tokens. - **Token Hierarchy**: - `last_q` and `exact_answer_last`/`after_last` share the highest attention in the final layers, implying the model prioritizes terminal input components for final output generation. - `exact_answer_before_first` and `exact_answer_first` show slightly earlier peaks (layers 24–28), possibly reflecting intermediate processing of answer boundaries. --- ### Interpretation This heatmap reveals how a transformer model allocates attention across input tokens at different processing depths: 1. **Early Layers (0–11)**: Likely handle low-level features (e.g., token embeddings, positional encoding) with minimal token-specific attention. 2. **Mid-Layers (12–22)**: Focus on answer-related tokens (`first_answer`, `second_answer`), suggesting these layers refine contextual relationships between question and answer components. 3. **Late Layers (24–30)**: Dominated by question and exact answer tokens, indicating these layers integrate high-level semantic understanding, particularly for terminal input elements. The concentration of high values in late layers for `last_q` and `exact_answer_last`/`after_last` suggests the model’s final output (e.g., generated answers) is heavily influenced by the last question and precise answer tokens. This aligns with transformer architectures, where deeper layers capture abstract, context-rich representations. **Notable Anomaly**: The `exact_answer_before_first` token shows elevated attention in layers 24–28 but declines sharply in layer 30, unlike other late-layer tokens. This could indicate a transitional role in answer boundary detection before final refinement in later layers. --- **Conclusion**: The heatmap demonstrates a clear progression of attention from low-level processing in early layers to high-level semantic integration in late layers, with terminal tokens (`last_q`, `exact_answer_last/after_last`) receiving the strongest focus in the model’s final stages.