license:
- cc-by-4.0
- apache-2.0
language:
- en
task_categories:
- text-classification
- question-answering
tags:
- rag
- evidence-sufficiency
- multi-hop-qa
- counterfactual
- benchmark
- chaincheck
pretty_name: ChainCheck
ChainCheck
Counterfactual controls for evaluating evidence-sufficiency scorers in retrieval-augmented QA. From ChainCheck: When Near-Perfect Evidence-Sufficiency Scores Fail to Distinguish Chain-Sensitive Systems (preprint forthcoming).
A sufficiency scorer can look near-perfect on an ordinary benchmark by reacting to any edit in the passages. ChainCheck adds a matched control so the two can be told apart:
| Cell | Chain | Edit |
|---|---|---|
| A | intact | none |
| B | intact | consistent bridge-entity swap |
| D | broken (later-hop substitution) | yes |
Nominal = AUC(A > D), chain effect = AUC(B > D) − 0.5, edit effect = AUC(A > B) − 0.5, chain selectivity Σ = CE − |EE|. These are matched contrasts, not an additive decomposition.
Files
| File | Rows | Source license |
|---|---|---|
musique_test.jsonl.gz |
1716 | MuSiQue (Trivedi et al., TACL 2022), CC BY 4.0 |
musique_confirmation.jsonl.gz |
678 | MuSiQue (Trivedi et al., TACL 2022), CC BY 4.0 |
twowiki_replication.jsonl.gz |
1392 | 2WikiMultiHopQA (Ho et al., COLING 2020), Apache-2.0 |
Each row is one (question, passage set) item: item_id, pair_id, variant (real, or fict for a synthetic replacement entity), cell (A/B/D), query, passages, answer, bridge, replacement, chain, edit, earlier_support, later_support, label_sufficient, cb_27b_correct (Qwen3.8-27B answered closed-book; the paper excludes these pairs) and source. Only quadruples that pass the Q1 to Q9 validator are included. Labels follow mechanically from source provenance; no human or LLM labels.
Evaluate your scorer
python chaincheck_eval.py --data twowiki_replication.jsonl.gz --scores my_scores.jsonl --exclude-cb27b
my_scores.jsonl holds one {"item_id": ..., "score": float} per item (higher = more sufficient). The script needs only numpy and reproduces the paper's reference-judge numbers exactly.
Licenses
Derived from MuSiQue (CC BY 4.0) and 2WikiMultiHopQA (Apache-2.0); passage text originates from Wikipedia (CC BY-SA). Each file keeps its source license; cite the original datasets alongside this one.
Models
Judges trained with these controls: ChainCheck-Judge collection.