Compare-Methods Blog Assets

This folder contains a blog-ready Chart.js bundle for the existing Llama patch detection figures generated from Compare-Methods.ipynb.

Scope:

• Datasets: Coco, Broden-OpenSurfaces, Broden-Pascal, iSarcasm, GoEmotions
• Model: Llama
• Sample type: patch
• Concept series: avg, linsep, kmeans, linsep kmeans
• Baselines: prompt, random
• Metric: weighted-average f1
• Percent-through-model: 100

Files:

• export_llama_patch_detection_vs_sparsity_bundle.py: Rebuilds the JSON bundle from raw files in Experiments/.
• llama_patch_detection_vs_sparsity.json: Exported chart data for the five blog charts.
• llama_patch_detection_vs_sparsity_chart.js: Native Chart.js renderer for the full bundle.
• llama_patch_detection_vs_sparsity_preview.html: Standalone preview page.

Suggested usage:

1. Copy llama_patch_detection_vs_sparsity.json and llama_patch_detection_vs_sparsity_chart.js into your blog assets.
2. Load Chart.js on the page.
3. Add a container like <div id="compare-methods-charts"></div>.
4. Fetch the JSON and call renderCompareMethodsCharts("#compare-methods-charts", bundle).
5. Serve the assets over HTTP for preview. The HTML file uses fetch, so file:// loading may be blocked by the browser.

Note:

• There is no existing Blog_Posts/ directory in this repo, so this bundle lives under Figs/Blog_Figs/detection_vs_sparsity/.
• The blog charts correspond to the paper figures saved as Figs/Paper_Figs/patch_*_Llama_dataset_detection.png.