mv1137/p2026-002-quantization-context-compression-results

Deployed Quantization Tier and Lossy Context Compression in Extractive QA

This result dataset mirrors the version-1.0.0 reproducibility artifact: 10.5281/zenodo.22847291. The versioned report and full replication sources are maintained together in the research-artifacts repository. Cite the exact Zenodo version for the frozen evidence; this Hugging Face copy is a discovery mirror.

• Matthew Schwartz — ORCID 0009-0009-4171-7247

This dataset is the aggregate-only evidence for "No Detected Interaction Between Deployed Quantization Tier and Lossy Context Compression in Extractive QA." The study asks whether lossy prompt compression costs more accuracy when a local deployment of one 27B model moves to a lower-footprint quantized artifact.

Headline result

Three deployments of Qwen3.8-27B on a 12 GiB laptop GPU answered the same extractive questions (120 WildChat, 117 private) under four conditions:

• Deployments: Q4_K_XL partly offloaded, EXL3 2.0 bpw, and IQ2_S.
• Conditions: the original text, LLMLingua-2, and two rungs of a stopword/code-stripping pipeline.

Findings:

• No interaction detected. None of the 12 paired difference-in-differences tests survives Holm correction. The estimates run from −5.0 to 6.0 percentage points, with a pointwise interval envelope of −11.6 to 12.9. This is a bounded non-detection, not an equivalence result.
• Compressor choice is the larger effect. LLMLingua-2 costs 15–21 points, and the naive pipeline costs 47–62 points.

Files

Path	Description	Rows
data/cells.csv	Per corpus, deployment, and condition: n, exact-correct count, accuracy with a Wilson interval, paired cost versus original with a Newcombe interval, exact McNemar p, mean request time, and correct answers per minute	24
data/interaction.csv	Per corpus, condition, and low deployment: the difference-in-differences estimate, CR2 interval and SE, Satterthwaite df, CR2 p, Holm p and decision, sign-flip p, and the number of source clusters	12
data/claims.json	The complete aggregate statistical and provenance record the tables are derived from	—

The tables contain aggregates only. No conversation text, question, answer, response, per-item row, or identifier from either corpus is included. The private corpus can never be released.

Reproduce the paper's numbers

python3 check_claims.py

The standard-library checker in the artifact recomputes every interval, test, and Holm decision from the per-cluster histograms in claims.json. It also re-derives both tables and every numeric macro printed in the paper.

Scope

• One dense 27B checkpoint, three artifact–engine–placement deployments, one laptop, extractive QA, and one sampled decode per cell.
• The harness requested seed 42, but the pinned TabbyAPI server did not apply it for EXL3. Q4 question generation through llama.cpp and item ordering were separately seeded. Exact analysis reproduction uses the saved outputs.
• A corrective exploratory rerun, not an independent confirmation.
• Engine, format, footprint, and GPU residency change together.

License and AI assistance

Original measurement records are offered under CC BY 4.0 to the extent the author holds rights; code is MIT. WildChat-1M (ODC-By) text is not redistributed.

Claude (Anthropic) and Codex (OpenAI) assisted with evaluation-harness development, experiment orchestration, evidence auditing, analysis and figure code, citation checking, and manuscript drafting and revision. Matthew Schwartz directed the research and is responsible for its methods, results, interpretation, claims, citations, rights, and released artifacts.