Update questions, pages, answers and document files

README.md

@@ -66,7 +66,7 @@ lists every document with its source URL, including those not hosted here. `page
 
 ## Variant notes
 
-Pages with `redistribution` other than `permitted` are excluded. XL-DocBench question text, gold answers and scoring rules are withheld (`text_withheld`), and `resolve_xl_docbench.py` fills question text and the benchmark answer by `id` from microsoft/XL-DocBench. Answers to questions on restricted documents and tool outputs that touch them read `[omitted from the public variant: source licence]`. Scores, call counts, seconds and tokens stay.
+Pages with `redistribution` other than `permitted` are excluded. XL-DocBench question text, gold answers and scoring rules are withheld (`text_withheld`), and `resolve_xl_docbench.py` fills question text and the benchmark answer by `id` from microsoft/XL-DocBench. The current release lacks five of these questions, four of them in `unseen`, and the script names them. Answers to questions on restricted documents and tool outputs that touch them read `[omitted from the public variant: source licence]`. Scores, call counts, seconds and tokens stay.
 
 ## Sources and attribution
 

answers/train.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:fc7c1e7b2e2b235c7481d07235df117bb2f525abeda129cf96e840e31da1e0c0
-size 254671
+oid sha256:c1842e56bffc8df0c8d36142debf245af5fef90b67049a5e14ca3e5b8e0d0b7f
+size 254213

resolve_xl_docbench.py

@@ -10,13 +10,16 @@ for name in ("qa_single_doc.jsonl", "qa_cross_doc.jsonl"):
     for line in urlopen(BASE + name):
         row = json.loads(line)
         source[row["question_id"]] = row
-for path_name in ("rag_misses", "multi_hop", "holdout", "holdout2"):
+for path_name in ("rag_misses", "multi_hop", "holdout", "holdout2", "unseen"):
     try:
         rows = parquet.read_table(f"questions/{path_name}.parquet").to_pylist()
     except FileNotFoundError:
         continue
+    missing = [row["id"] for row in rows if row["text_withheld"] and row["id"] not in source]
+    if missing:
+        print(f"{path_name}: not in the current XL-DocBench release: {', '.join(missing)}")
     for row in rows:
-        if row["text_withheld"]:
+        if row["text_withheld"] and row["id"] in source:
             row["question"], row["gold"] = source[row["id"]]["question"], str(source[row["id"]]["answer"]["value"])
     with open(f"questions_{path_name}_resolved.jsonl", "w") as handle:
         handle.writelines(json.dumps(row, ensure_ascii=False) + "\n" for row in rows)