Add files using upload-large-folder tool

MANIFEST.json

@@ -0,0 +1,47 @@
+{
+  "name": "ChainCheck v2",
+  "files": {
+    "musique_test.jsonl.gz": {
+      "rows": 1716,
+      "license": "MuSiQue (Trivedi et al., TACL 2022), CC BY 4.0",
+      "sha256": "44858e1641dd6ec846560a5791e01e1bba17f97d9f65052e1de38fa7f2d5b42c"
+    },
+    "musique_confirmation.jsonl.gz": {
+      "rows": 678,
+      "license": "MuSiQue (Trivedi et al., TACL 2022), CC BY 4.0",
+      "sha256": "7fcdd95036e8a0677f940a19d7580e58f8a85159c00fff5fbdc7845e2127aa8c"
+    },
+    "twowiki_replication.jsonl.gz": {
+      "rows": 1392,
+      "license": "2WikiMultiHopQA (Ho et al., COLING 2020), Apache-2.0",
+      "sha256": "be4e61202aa0a9fd4ee83dd199910263a56e145c2c3f652234b2ca95b584e284"
+    }
+  },
+  "counts": {
+    "musique_test": {
+      "fict|invalid": 36,
+      "fict|valid": 234,
+      "real|invalid": 125,
+      "real|valid": 338
+    },
+    "musique_confirmation": {
+      "fict|invalid": 14,
+      "fict|valid": 100,
+      "real|invalid": 40,
+      "real|valid": 126
+    },
+    "twowiki_replication": {
+      "fict|invalid": 15,
+      "fict|valid": 149,
+      "real|invalid": 170,
+      "real|valid": 315
+    }
+  },
+  "cells": {
+    "A": "chain alive, no edit",
+    "B": "chain alive, edit (consistent bridge swap)",
+    "D": "chain broken, edit (bridge substitution in later hop)"
+  },
+  "metrics": "Nominal=AUC(A>D) · Chain effect=AUC(B>D)-0.5 · Edit effect=AUC(A>B)-0.5 · Sigma=CE-|EE|",
+  "note": "Labels follow mechanically from dataset provenance and the Q1-Q9 validator; no human or LLM labels. Passage text originates from Wikipedia."
+}

README.md

@@ -0,0 +1,47 @@
+---
+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
+
+```bash
+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](https://huggingface.co/ThakiCloud).

chaincheck_eval.py

@@ -0,0 +1,71 @@
+#!/usr/bin/env python3
+"""ChainCheck evaluator (standalone; needs only numpy).
+
+Score every item of a ChainCheck file with your evidence-sufficiency system (higher = "the passages are
+sufficient"), then:
+
+    python chaincheck_eval.py --data twowiki_replication.jsonl.gz --scores my_scores.jsonl [--exclude-cb27b]
+
+my_scores.jsonl: one JSON object per line, {"item_id": ..., "score": float}.
+
+Reports, separately for real and synthetic replacement entities:
+  Nominal = AUC(A > D)            what an ordinary benchmark reports
+  CE      = AUC(B > D) - 0.5      chain effect: response to breaking the chain, edit held fixed
+  EE      = AUC(A > B) - 0.5      edit effect: response to an edit that leaves the chain intact
+  Sigma   = CE - |EE|             chain selectivity; chain-selective if the 95% lower bound > 0
+Pair-bootstrap 95% intervals (2,000 resamples, seed 0). These are matched contrasts, not an additive
+decomposition: Nominal is not CE + EE.
+--exclude-cb27b drops pairs whose question Qwen3.8-27B answered closed-book (the paper's closed-book-hard filter).
+"""
+import argparse, gzip, json
+from collections import defaultdict
+import numpy as np
+
+
+def load(path):
+    op = gzip.open if path.endswith(".gz") else open
+    with op(path, "rt", encoding="utf-8") as f:
+        return [json.loads(x) for x in f if x.strip()]
+
+
+def wins(s, hi, lo):
+    return np.array([(s[h] > s[l]) + 0.5 * (s[h] == s[l]) for h, l in zip(hi, lo)], dtype=float)
+
+
+def metrics(s, quads, n_boot=2000, seed=0):
+    A = [q["A"] for q in quads]; B = [q["B"] for q in quads]; D = [q["D"] for q in quads]
+    W = {"nominal": wins(s, A, D), "chain": wins(s, B, D), "edit": wins(s, A, B)}
+    idx = np.random.default_rng(seed).integers(0, len(quads), size=(n_boot, len(quads)))
+    ce_b = W["chain"][idx].mean(1) - 0.5; ee_b = W["edit"][idx].mean(1) - 0.5; sig_b = ce_b - np.abs(ee_b)
+    ci = lambda x: [round(float(np.percentile(x, 2.5)), 4), round(float(np.percentile(x, 97.5)), 4)]
+    ce, ee = W["chain"].mean() - 0.5, W["edit"].mean() - 0.5
+    return {"n_pairs": len(quads), "nominal": round(float(W["nominal"].mean()), 4), "CE": round(float(ce), 4),
+            "CE_ci": ci(ce_b), "EE": round(float(ee), 4), "EE_ci": ci(ee_b), "sigma": round(float(ce - abs(ee)), 4),
+            "sigma_ci": ci(sig_b), "chain_selective": bool(np.percentile(sig_b, 2.5) > 0)}
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--data", required=True); ap.add_argument("--scores", required=True)
+    ap.add_argument("--exclude-cb27b", action="store_true")
+    a = ap.parse_args()
+    rows = load(a.data)
+    sc = {x["item_id"]: float(x["score"]) for x in load(a.scores)}
+    missing = [r["item_id"] for r in rows if r["item_id"] not in sc]
+    if missing:
+        raise SystemExit(f"{len(missing)} items have no score, e.g. {missing[:3]}")
+    quads = defaultdict(dict)
+    for r in rows:
+        if a.exclude_cb27b and r["cb_27b_correct"]:
+            continue
+        quads[(r["pair_id"], r["variant"])][r["cell"]] = r["item_id"]
+    out = {}
+    for variant in ("real", "fict"):
+        qs = [q for (p, v), q in sorted(quads.items()) if v == variant and len(q) == 3]
+        if qs:
+            out[variant] = metrics(sc, qs)
+    print(json.dumps(out, indent=1))
+
+
+if __name__ == "__main__":
+    main()

musique_confirmation.jsonl.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7fcdd95036e8a0677f940a19d7580e58f8a85159c00fff5fbdc7845e2127aa8c
+size 339194

musique_test.jsonl.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:44858e1641dd6ec846560a5791e01e1bba17f97d9f65052e1de38fa7f2d5b42c
+size 826824

twowiki_replication.jsonl.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be4e61202aa0a9fd4ee83dd199910263a56e145c2c3f652234b2ca95b584e284
+size 507902