Add the build script; card says how to rebuild

README.md

@@ -32,4 +32,5 @@ loaded as one dataset directly.
 
 Licenses are those of the underlying datasets; see the source dataset's
 [manifest](https://github.com/uspraveen/Jevify/blob/main/docs/DATASETS.md). Rows are sampled with a fixed
-seed by `task_diagnostics/build_jevbench.py`.
+seed by `build_jevbench.py` in this repo:
+`python build_jevbench.py --out DIR [--push tasksource/jevbench]` rebuilds it exactly.

build_jevbench.py

@@ -0,0 +1,54 @@
+"""Recast Praveenrajus/jev-bench into one flat dataset: N rows per data config (model-prediction dumps excluded),
+merged, the config as a `task` column, every field a string (soft_label "" when absent). The per-config files
+disagree on soft_label's type, so the raw configs cannot be loaded as one dataset.
+
+    python build_jevbench.py --out DIR [--push tasksource/jevbench]   (or python -m task_diagnostics.build_jevbench)
+"""
+import argparse, json, random
+from pathlib import Path
+
+SRC, REV = "Praveenrajus/jev-bench", "c37b0f6ab1687376ec9ae8dbd9f343163a68252b"
+CONFIGS = ("arc_challenge banking77 boolq chaosnli civil_comments clinc150 fever_evidence go_emotions "
+           "helpsteer2_helpfulness helpsteer2_verbosity ledgar massive measuring_hate_speech mmlu mnli paws "
+           "sms_spam sst5 strategyqa_closed strategyqa_grounded stsb yelp5").split()
+
+
+def build(n, seed, splits):
+    from huggingface_hub import hf_hub_download
+    out = {}
+    for split in splits:
+        rows = []
+        for c in CONFIGS:
+            try:
+                p = hf_hub_download(SRC, f"data/{c}/{split}.jsonl", repo_type="dataset", revision=REV)
+            except Exception:
+                continue   # chaosnli has a test split only
+            src = sorted((json.loads(l) for l in open(p)), key=lambda r: r["id"])
+            random.Random(f"{seed}/{c}/{split}").shuffle(src)
+            rows += [{"task": c, **{k: "" if r.get(k) is None else str(r[k]) for k in
+                      ("id", "primitive", "state", "question", "label", "soft_label", "meta")}} for r in src[:n]]
+        out[split] = rows
+        print(f"[jevbench] {split}: {len(rows)} rows over {len({r['task'] for r in rows})} tasks", flush=True)
+    return out
+
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__.splitlines()[0])
+    ap.add_argument("--out", required=True)
+    ap.add_argument("--n", type=int, default=200, help="rows per config and split")
+    ap.add_argument("--splits", default="test,validation")
+    ap.add_argument("--seed", type=int, default=0)
+    ap.add_argument("--push", help="hub repo id to push to")
+    a = ap.parse_args()
+    data = build(a.n, a.seed, a.splits.split(","))
+    Path(a.out).mkdir(parents=True, exist_ok=True)
+    for split, rows in data.items():
+        (Path(a.out) / f"{split}.jsonl").write_text("".join(json.dumps(r) + "\n" for r in rows))
+    if a.push:
+        import datasets
+        dd = datasets.DatasetDict({s: datasets.Dataset.from_list(r) for s, r in data.items()})
+        dd.push_to_hub(a.push, commit_message=f"{a.n}/config sample of {SRC}@{REV[:7]}, seed {a.seed}")
+
+
+if __name__ == "__main__":
+    main()