From c25e9cb084bfb6e8f19a0ef86b2229769cb7c8e0 Mon Sep 17 00:00:00 2001
From: Jake Poznanski <jakep@allenai.org>
Date: Wed, 19 Mar 2025 18:57:00 +0000
Subject: [PATCH] Addxing some fixes

---
 olmocr/bench/benchmark.py  |  1 +
 olmocr/bench/test_tests.py | 19 +++++++++++++++++++
 2 files changed, 20 insertions(+)

diff --git a/olmocr/bench/benchmark.py b/olmocr/bench/benchmark.py
index 9fd1737..1abc3dc 100644
--- a/olmocr/bench/benchmark.py
+++ b/olmocr/bench/benchmark.py
@@ -306,6 +306,7 @@ def main():
             half_width = ((ci[1] - ci[0]) / 2) * 100
             ciw_str = f"± {half_width:0.1f}%"
         print(f"{candidate_name:20s} : Average Score: {status} {ciw_str}")
+        test_type_breakdown.sort(key=lambda f: f[0])
         for ttype, scores in test_type_breakdown.items():
             avg = sum(scores) / len(scores) * 100 if scores else 0.0
             print(f"    {ttype:8s}: {avg:0.1f}% average pass rate over {len(scores)} tests")
diff --git a/olmocr/bench/test_tests.py b/olmocr/bench/test_tests.py
index 89993ae..61f59b8 100644
--- a/olmocr/bench/test_tests.py
+++ b/olmocr/bench/test_tests.py
@@ -572,6 +572,25 @@ Some text before the table.
         result, explanation = test.run(table)
         self.assertTrue(result, explanation)
 
+    def test_big_table(self):
+        # TODO this would work with html tables with a rowspan?
+        table = """| Question - – Satisfaction on scale of 10 | Response | Resident Sample | Business Sample |
+|----------------------------------------|----------|----------------|-----------------|
+| Planning for and managing residential, commercial and industrial development | Rating of 8, 9 or 10 | 13% | 11% |
+| | Average rating | 6.4 | 5.7 |
+| | Don’t know responses | 11% | 6% |
+| Environmental protection, support for green projects (e.g. green grants, building retrofits programs, zero waste) | Rating of 8, 9 or 10 | 35% | 34% |
+| | Average rating | 8.0 | 7.5 |
+| | Don’t know responses | 8% | 6% |
+| Providing and maintaining parks and green spaces | Rating of 8, 9 or 10 | 42% | 41% |
+| | Average rating | 7.7 | 7.3 |
+| | Don’t know responses | 1% | 1% |
+
+Base: Resident respondents (n=1,315) and Business respondents (n=397)
+"""
+        test = TableTest(pdf="test.pdf", page=1, id="test_id", type=TestType.TABLE.value, cell="Planning for and managing residential, commercial and industrial development", down="Environmental protection,\nsupport for green projects\n(e.g. green grants,\nbuilding retrofits programs,\nzero waste)")
+        result, explanation = test.run(table)
+        self.assertTrue(result, explanation)
 
 class TestBaselineTest(unittest.TestCase):
     """Test the BaselineTest class"""