v1.0.0 — Context Curation Experiment

Context Curation Improves LLM Answer Quality: An Empirical Evaluation on SQuAD

Initial release containing the complete experiment, results, and paper.

Key Results (200 SQuAD 2.0 questions)

• Exact Match: 55.0% → 60.5% (+5.5 pp) with curation
• F1: 76.1% → 76.9% (+0.8 pp) with curation
• Token reduction: 92.0% (1,191 → 95 avg prompt tokens)

Contents

• experiment.py — Complete experiment script
• paper.pdf — Paper (PDF)
• paper.txt — Paper (plain text)
• paper.tex — LaTeX source
• results.jsonl — Per-question results
• summary.json — Aggregate statistics

Hardware

M1 Mac Mini, 16GB RAM, 512GB SSD. All inference local via Ollama.

Models

• Answer: Llama 3.1 8B
• Curator: Qwen 2.5 7B