Gespeichert in:
| 1. Verfasser: | |
|---|---|
| Format: | Recurso digital |
| Sprache: | |
| Veröffentlicht: |
Zenodo
2026
|
| Online-Zugang: | https://doi.org/10.5281/zenodo.18626830 |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
Inhaltsangabe:
- <h2>Context Curation Improves LLM Answer Quality: An Empirical Evaluation on SQuAD</h2> <p>Initial release containing the complete experiment, results, and paper.</p> <h3>Key Results (200 SQuAD 2.0 questions)</h3> <ul> <li><strong>Exact Match</strong>: 55.0% → 60.5% (+5.5 pp) with curation</li> <li><strong>F1</strong>: 76.1% → 76.9% (+0.8 pp) with curation</li> <li><strong>Token reduction</strong>: 92.0% (1,191 → 95 avg prompt tokens)</li> </ul> <h3>Contents</h3> <ul> <li><code>experiment.py</code> — Complete experiment script</li> <li><code>paper.pdf</code> — Paper (PDF)</li> <li><code>paper.txt</code> — Paper (plain text)</li> <li><code>paper.tex</code> — LaTeX source</li> <li><code>results.jsonl</code> — Per-question results</li> <li><code>summary.json</code> — Aggregate statistics</li> </ul> <h3>Hardware</h3> <p>M1 Mac Mini, 16GB RAM, 512GB SSD. All inference local via Ollama.</p> <h3>Models</h3> <ul> <li>Answer: Llama 3.1 8B</li> <li>Curator: Qwen 2.5 7B</li> </ul>