STALI Paper 2 — cross-dataset adapters

Public multi-seed LoRA checkpoints for A Late-Interaction Retriever Matches Graph-RAG on Multi-Hop Document Retrieval: A Matched-Reader Re-Evaluation.

Canonical code and data release: https://github.com/apertis-ai/multihop-doc-qa/tree/main/paper2

a2_seed{1,2,3}/: DocHop-QA A2 runs.
dochop_multiseed/{s1,s2,s2_pure}_seed{1,2,3}/: additional DocHop-QA trained-system adapters, run metadata, and evaluations.
dochop_eval_results/: retained per-query outputs for the remaining DocHop-QA baselines and context-length runs.
dochop_eval_corrected/: corrected top-5 section IDs and metrics for all 25 retained outputs that contain per-query top-5 IDs. Historical source files are preserved unchanged.
e6_multihoprag_seed{1,2,3,4,42}/: MultiHop-RAG runs.
e6_hotpotqa_seed{1,2,3}/: HotpotQA runs.

Each run contains a canonical adapter/ directory and retained training checkpoints. Load the adapter/ subdirectory with PEFT on top of lightonai/GTE-ModernColBERT-v1. Per-query evaluations associated with the runs are included where available; the GitHub release contains the compact consolidated result package and integrity manifest.

The adapters target ModernBERT attention modules Wqkv and Wo with LoRA rank 16 and alpha 32.

The preserved S2-pure metadata records 6,343 training samples. This corrects the paper text's 6,310 count; see the canonical release for the full cohort and historical RNG notes.

The corrected DocHop-QA files repair a legacy /N/A suffix mismatch and embed both historical and corrected section R@5 values plus the number of affected queries. Top-10 section metrics remain historical because top-10 IDs were not retained. Three metric-only outputs (e1_bge_m3, e1_qwen3_emb_4b, and e2_rerank_bge_v2m3) cannot be corrected without rerunning retrieval. The S3 context-labelled outputs are also retained as historical records: the old runners did not apply the YAML token limits, so they are not evidence for a 512–4096-token ablation.