From 0b5f71ae4701cbfcfc672deeb36ecdd503da3a18 Mon Sep 17 00:00:00 2001
From: Yiorgis Gozadinos <ggozadinos@gmail.com>
Date: Wed, 3 Jun 2026 16:03:42 +0300
Subject: [PATCH] exclude corrupt mi_phone.pdf from MMLongBench-Doc

---
 .../evaluations/datasets/mmlongbench.py       | 10 ++++-
 evaluations/tests/test_datasets.py            | 42 +++++++++++++++++++
 2 files changed, 51 insertions(+), 1 deletion(-)

diff --git a/evaluations/evaluations/datasets/mmlongbench.py b/evaluations/evaluations/datasets/mmlongbench.py
index 436f9e1f..87dcd71c 100644
--- a/evaluations/evaluations/datasets/mmlongbench.py
+++ b/evaluations/evaluations/datasets/mmlongbench.py
@@ -17,6 +17,13 @@ REPO_ID = "yubo2333/MMLongBench-Doc"
 PDF_SUBDIR = "documents"
 _LIST_FIELDS = ("evidence_pages", "evidence_sources")
 
+# The HF repo serves a blob for these files whose content does not match the
+# document the questions were written against (the LFS pointer and the served
+# bytes diverge, and the bytes are an unrelated PDF). They are unrecoverable
+# upstream, so the document and all its questions are dropped to keep the
+# benchmark answerable and reproducible.
+_EXCLUDED_DOCS = frozenset({"mi_phone.pdf"})
+
 
 def get_cache_dir() -> Path:
     cache_dir = Path.home() / ".cache" / "haiku.rag" / "evaluations" / "mmlongbench"
@@ -30,6 +37,7 @@ def ensure_pdfs_downloaded() -> Path:
         repo_id=REPO_ID,
         repo_type="dataset",
         allow_patterns=f"{PDF_SUBDIR}/*.pdf",
+        ignore_patterns=[f"{PDF_SUBDIR}/{name}" for name in _EXCLUDED_DOCS],
         local_dir=cache_dir,
     )
     return cache_dir / PDF_SUBDIR
@@ -55,7 +63,7 @@ def load_qa_records() -> list[dict[str, Any]]:
     global _qa_records
     if _qa_records is not None:
         return _qa_records
-    rows = _load_hf_qa_split()
+    rows = [r for r in _load_hf_qa_split() if r.get("doc_id") not in _EXCLUDED_DOCS]
     for row in rows:
         for field in _LIST_FIELDS:
             row[field] = _parse_list_field(row.get(field))
diff --git a/evaluations/tests/test_datasets.py b/evaluations/tests/test_datasets.py
index 66f9a844..bfa86fcc 100644
--- a/evaluations/tests/test_datasets.py
+++ b/evaluations/tests/test_datasets.py
@@ -271,13 +271,55 @@ class TestMMLongBenchDoc:
             },
         ]
 
+        import evaluations.datasets.mmlongbench as m
+
+        m._qa_records = None
         with patch(
             "evaluations.datasets.mmlongbench._load_hf_qa_split",
             return_value=raw_rows,
         ):
             records = load_qa_records()
+        m._qa_records = None
 
         assert records[0]["evidence_pages"] == [3, 5]
         assert records[0]["evidence_sources"] == ["Table", "Pure-text"]
         assert records[1]["evidence_pages"] == []
         assert records[1]["evidence_sources"] == []
+
+    def test_load_qa_records_drops_excluded_docs(self) -> None:
+        from unittest.mock import patch
+
+        import evaluations.datasets.mmlongbench as m
+
+        raw_rows = [
+            {
+                "doc_id": "mi_phone.pdf",
+                "doc_type": "Guidebook",
+                "question": "Q?",
+                "answer": "A",
+                "evidence_pages": "[1]",
+                "evidence_sources": "['Pure-text']",
+                "answer_format": "Str",
+            },
+            {
+                "doc_id": "keep.pdf",
+                "doc_type": "Brochure",
+                "question": "Q2?",
+                "answer": "A2",
+                "evidence_pages": "[2]",
+                "evidence_sources": "['Table']",
+                "answer_format": "Str",
+            },
+        ]
+
+        m._qa_records = None
+        with patch(
+            "evaluations.datasets.mmlongbench._load_hf_qa_split",
+            return_value=raw_rows,
+        ):
+            records = load_qa_records()
+        m._qa_records = None
+
+        doc_ids = {r["doc_id"] for r in records}
+        assert "mi_phone.pdf" not in doc_ids
+        assert "keep.pdf" in doc_ids