Doc2LoRA idea genes: reproduction code and archived intermediates for "Doc2LoRA Provides Decodable Representations of Scientific Ideas"

Abstract

The reproduction workflow and the archived intermediate results for the paper. Everything needed to rebuild the manuscript's tables is here: the code is a snapshot of the trimmed reproduction repository, and the bundles hold the computed intermediates, so a reader does not have to re-derive them. Reproducing from the raw corpora instead needs roughly 200 GB of intermediates and several GPU-days. doc2lora-embedding-code- .tar.gz — the workflow itself: a Snakemake pipeline trimmed to the dependency closure of what the manuscript reads (24 rule files; snakemake -n paper_assets resolves 330 jobs from the raw corpora). Includes REPRODUCE.md, which maps every figure, table and quoted number to the rule that produces it, and data/ARTIFACTS.tsv, a SHA-256 per archived file. doc2lora-results.tar.zst (1227 files) — everything the manuscript's tables and figures are computed from that needs a GPU, the licensed APS text, or an LLM judge to produce: the paired per-unit score pools behind the bootstrap confidence intervals, the invertible adapter weights, every method's raw cluster labels and the judge verdicts, the pair-axis decodes (Doc2LoRA and ICAE), the PACS node set, and the manuscript's own tables for comparison. With this bundle alone, snakemake paper_assets --rerun-triggers mtime rebuilds every table and figure the paper reads, running only CPU rules. doc2lora-s2and.tar.zst (50 files) — the gene and text embeddings for the five author-name-disambiguation benchmarks (zbMATH, QIAN, ArnetMiner, PubMed, KISTI), so the disambiguation rows can be re-scored from the vectors up. Unpack the code snapshot, then fetch and verify the bundles against data/ARTIFACTS.tsv (SHA-256 per file): tar xzf doc2lora-embedding-code- .tar.gz && cd doc2lora-embedding- cp workflow/config.template.yaml workflow/config.yaml # edit the paths python scripts/fetch_artifacts.py results --record snakemake paper_assets -j4 --rerun-triggers mtime # rebuilds every table, CPU only The 644k-paper APS gene matrices (~137 GB) are not included: they exceed a Zenodo record and are a deterministic function of the APS corpus plus the published hypernetwork checkpoints (snakemake all_embeddings). The APS corpus itself is licensed and is not redistributed here.