# Technical Document Extraction ## Figure (a): Line Chart ### Title - "Our method needs a smaller calibration set" ### Axes - **X-axis**: - Label: `# calibration sequences (×2048 tokens)` - Values: 8, 16, 32, 64, 128, 192, 256 - **Y-axis**: - Label: `Perplexity` - Range: 13 to 14 ### Legend - **GPTQ**: Green diamonds (`□`) - **Ours**: Orange triangles (`▲`) ### Data Trends - **GPTQ**: - Starts at 13.9 perplexity at 8 sequences - Decreases to ~13.1 by 256 sequences - **Ours**: - Starts at 13.3 perplexity at 8 sequences - Decreases to ~13.1 by 256 sequences - **Observation**: Both methods show decreasing perplexity with more calibration sequences, but "Ours" maintains lower perplexity across all sequence counts. --- ## Figure (b): Table ### Title - "Our method is more robust to calibration set distribution" ### Table Structure | | **GPTQ** | **Ours** | |----------------|----------------|----------------| | **Eval** | | | | **Calib** | | | | **PubMed** | 32.48 | 32.56 | | **Enron** | 50.41 | 45.07 | | **PubMed** | 34.81 | 33.16 | | **Enron** | 45.52 | 44.57 | ### Arrows & Differences - **GPTQ PubMed → Ours PubMed**: `+0.08` (32.48 → 32.56) - **GPTQ Enron → Ours Enron**: `+4.89` (50.41 → 45.07) - **PubMed Enron (GPTQ) → Ours Enron**: `+0.60` (45.52 → 44.57) - **PubMed Enron (Ours) → GPTQ Enron**: `+0.50` (44.57 → 45.07) ### Key Observations - "Ours" method shows smaller performance gaps across evaluation sets compared to GPTQ. - Largest improvement observed in Enron dataset (`+4.89` reduction). - Consistent robustness across both PubMed and Enron evaluation sets.