Datasets:

Netzerep
/

monitors-evaluation-awareness

# One-off patch of the dataset's v1.0 shards to v1.1 (lead, 2026-11-05): conclusion columns in the readout's kinds
# (target_score + grader_rules -> grader; hedge folded by build_dataset.conclusion) and the prefill "hedge" condition
# dropped. Same shards, schema and compression; rewrites build_report.json counts.
# Usage: python patch_v1_1.py <dataset folder>
import collections, glob, json, math, os, sys
import pyarrow as pa, pyarrow.compute as pc, pyarrow.parquet as pq

H = sys.argv[1]
src = open(f"{H}/scripts/build_dataset.py").read()
exec(src[src.index("CONCL_MERGE = {"):src.index("\n\n\n# ----", src.index("CONCL_MERGE = {"))])  # CONCL_MERGE, conclusion()
rep = json.load(open(f"{H}/build_report.json"))
cnt = dict(rows=0, setting=collections.Counter(), bench=collections.Counter(), mon=collections.Counter(), js=collections.Counter())
concl = collections.Counter(); dropped = 0
for f in sorted(glob.glob(f"{H}/data/reviews-*.parquet")):
    pf = pq.ParquetFile(f); t = pf.read(); n0 = t.num_rows
    rg = math.ceil(n0 / pf.metadata.num_row_groups)
    keep = pc.invert(pc.and_(pc.equal(t["setting"], "prefill"), pc.fill_null(pc.equal(t["prefill_condition"], "hedge"), False)))
    t = t.filter(keep); dropped += n0 - t.num_rows
    mains, alsos = t["vea_conclusion"].to_pylist(), t["vea_conclusions_also"].to_pylist()
    new = [conclusion({"main": m, "also": a}) if m is not None else (None, a) for m, a in zip(mains, alsos)]
    i = t.schema.get_field_index("vea_conclusion"); t = t.set_column(i, t.schema.field(i), pa.array([x[0] for x in new], t.schema.field(i).type))
    i = t.schema.get_field_index("vea_conclusions_also"); t = t.set_column(i, t.schema.field(i), pa.array([x[1] for x in new], t.schema.field(i).type))
    pq.write_table(t, f, compression="zstd", row_group_size=rg)
    cnt["rows"] += t.num_rows
    for k, col in (("setting", "setting"), ("bench", "benchmark"), ("mon", "monitor_model")):
        cnt[k].update(t[col].to_pylist())
    cnt["js"].update(f"{s}/{j}" for s, j in zip(t["setting"].to_pylist(), t["judge_status"].to_pylist()))
    concl.update(x[0] for x in new if x[0])
rep.update(rows=cnt["rows"], rows_by_setting=dict(sorted(cnt["setting"].items())), rows_by_benchmark=dict(sorted(cnt["bench"].items())),
           rows_by_monitor=dict(sorted(cnt["mon"].items())), judge_status=dict(sorted(cnt["js"].items())),
           bytes=sum(os.path.getsize(f) for f in glob.glob(f"{H}/data/reviews-*.parquet")),
           v1_1_patch="2026-11-05: conclusion columns in the readout's kinds; prefill condition 'hedge' dropped (%d rows)" % dropped)
json.dump(rep, open(f"{H}/build_report.json", "w"), indent=1, sort_keys=True)
print("dropped", dropped, "rows", cnt["rows"]); print(dict(concl)); print(json.dumps({k: rep[k] for k in ("rows_by_setting", "rows_by_monitor", "judge_status")}, indent=1))

	# One-off patch of the dataset's v1.0 shards to v1.1 (lead, 2026-11-05): conclusion columns in the readout's kinds
	# (target_score + grader_rules -> grader; hedge folded by build_dataset.conclusion) and the prefill "hedge" condition
	# dropped. Same shards, schema and compression; rewrites build_report.json counts.
	# Usage: python patch_v1_1.py <dataset folder>
	import collections, glob, json, math, os, sys
	import pyarrow as pa, pyarrow.compute as pc, pyarrow.parquet as pq

	H = sys.argv[1]
	src = open(f"{H}/scripts/build_dataset.py").read()
	exec(src[src.index("CONCL_MERGE = {"):src.index("\n\n\n# ----", src.index("CONCL_MERGE = {"))]) # CONCL_MERGE, conclusion()
	rep = json.load(open(f"{H}/build_report.json"))
	cnt = dict(rows=0, setting=collections.Counter(), bench=collections.Counter(), mon=collections.Counter(), js=collections.Counter())
	concl = collections.Counter(); dropped = 0
	for f in sorted(glob.glob(f"{H}/data/reviews-*.parquet")):
	pf = pq.ParquetFile(f); t = pf.read(); n0 = t.num_rows
	rg = math.ceil(n0 / pf.metadata.num_row_groups)
	keep = pc.invert(pc.and_(pc.equal(t["setting"], "prefill"), pc.fill_null(pc.equal(t["prefill_condition"], "hedge"), False)))
	t = t.filter(keep); dropped += n0 - t.num_rows
	mains, alsos = t["vea_conclusion"].to_pylist(), t["vea_conclusions_also"].to_pylist()
	new = [conclusion({"main": m, "also": a}) if m is not None else (None, a) for m, a in zip(mains, alsos)]
	i = t.schema.get_field_index("vea_conclusion"); t = t.set_column(i, t.schema.field(i), pa.array([x[0] for x in new], t.schema.field(i).type))
	i = t.schema.get_field_index("vea_conclusions_also"); t = t.set_column(i, t.schema.field(i), pa.array([x[1] for x in new], t.schema.field(i).type))
	pq.write_table(t, f, compression="zstd", row_group_size=rg)
	cnt["rows"] += t.num_rows
	for k, col in (("setting", "setting"), ("bench", "benchmark"), ("mon", "monitor_model")):
	cnt[k].update(t[col].to_pylist())
	cnt["js"].update(f"{s}/{j}" for s, j in zip(t["setting"].to_pylist(), t["judge_status"].to_pylist()))
	concl.update(x[0] for x in new if x[0])
	rep.update(rows=cnt["rows"], rows_by_setting=dict(sorted(cnt["setting"].items())), rows_by_benchmark=dict(sorted(cnt["bench"].items())),
	rows_by_monitor=dict(sorted(cnt["mon"].items())), judge_status=dict(sorted(cnt["js"].items())),
	bytes=sum(os.path.getsize(f) for f in glob.glob(f"{H}/data/reviews-*.parquet")),
	v1_1_patch="2026-11-05: conclusion columns in the readout's kinds; prefill condition 'hedge' dropped (%d rows)" % dropped)
	json.dump(rep, open(f"{H}/build_report.json", "w"), indent=1, sort_keys=True)
	print("dropped", dropped, "rows", cnt["rows"]); print(dict(concl)); print(json.dumps({k: rep[k] for k in ("rows_by_setting", "rows_by_monitor", "judge_status")}, indent=1))