artifacts/paper_main/MANIFEST.md

Paper Artifact Manifest

This directory points to the reproducible artifacts for the current paper draft.

Working title: When Does Lexical Search Help Late-Interaction Retrieval?

Canonical Reproduction Entry Points

• Config: experiments/configs/paper_main.yaml
• End-to-end runner: experiments/scripts/run_paper_main.py
• Table compiler: experiments/scripts/compile_paper_tables.py
• Command sheet: artifacts/paper_main/COMMANDS.md
• Current result snapshot: artifacts/paper_main/RESULTS_SNAPSHOT.md

Run the full main experiment:

uv sync
uv run python experiments/scripts/prepare_paper_data.py --config experiments/configs/paper_main.yaml
uv run python experiments/scripts/run_paper_main.py --config experiments/configs/paper_main.yaml

The runner is resumable: it skips completed BM25, dense, fusion, rerank, and feature outputs unless --force is passed.

Current Completed Outputs Used For Analysis

The currently completed experiment outputs live under the historical experiment id paper8_chunked:

• Main analysis: results/paper8_chunked/analysis.md
• Paper table: results/paper8_chunked/paper_tables.md
• Paper table JSON: results/paper8_chunked/paper_tables.json
• Per-dataset metrics: results/paper8_chunked/<dataset>/*.json
• TREC run files: runs/paper8_chunked/<dataset>/*.run

The canonical paper_main.yaml includes the 7 datasets with complete BM25, dense, RRF, BM25-rerank, and RRF-rerank results:

• scifact
• nfcorpus
• arguana
• scidocs
• fiqa
• cqadupstack/android
• cqadupstack/mathematica

Reproducibility Notes

• Raw BEIR data and generated indexes are intentionally not tracked in version control.
• data/ is recreated by download_beir.py for standard BEIR datasets.
• CQADupStack subdomains are exported with export_hf_cqadupstack_to_beir.py.
• Runtime depends heavily on hardware. Quality metrics are the intended reproducibility target; local latency should be treated as diagnostic.
• Apple MPS can produce non-finite embeddings for some long/chunked inputs. The reranking script repairs affected rows by CPU re-encoding before scoring.