import random import unittest from olmocr.synth.augmentations import ( _apply_typo, _has_skip_ancestor, introduce_text_errors, ) class TestApplyTypo(unittest.TestCase): """Unit tests for the _apply_typo helper.""" def test_preserves_first_and_last_char(self): rng = random.Random(0) for seed in range(50): rng.seed(seed) typo = _apply_typo("abcdef", rng) self.assertEqual(typo[0], "a") self.assertEqual(typo[-1], "f") def test_swap_strategy(self): # Force swap by controlling the RNG rng = random.Random() word = "abcdef" found_swap = False for seed in range(200): rng.seed(seed) typo = _apply_typo(word, rng) # A swap keeps same length and same set of chars if len(typo) == len(word) and sorted(typo) == sorted(word) and typo != word: found_swap = True break self.assertTrue(found_swap, "Expected to find at least one swap typo") def test_delete_strategy(self): rng = random.Random() word = "abcdef" found_delete = False for seed in range(200): rng.seed(seed) typo = _apply_typo(word, rng) if len(typo) == len(word) - 1: found_delete = True break self.assertTrue(found_delete, "Expected to find at least one delete typo") def test_duplicate_strategy(self): rng = random.Random() word = "abcdef" found_dup = False for seed in range(200): rng.seed(seed) typo = _apply_typo(word, rng) if len(typo) == len(word) + 1: found_dup = True break self.assertTrue(found_dup, "Expected to find at least one duplicate typo") def test_short_word_unchanged(self): rng = random.Random(42) # 2-char word has 0 interior chars, cannot be mutated self.assertEqual(_apply_typo("ab", rng), "ab") def test_three_char_word(self): # 3-char word has exactly 1 interior char; swap requires >=2 interior, so only delete/dup rng = random.Random() word = "abc" for seed in range(50): rng.seed(seed) typo = _apply_typo(word, rng) self.assertEqual(typo[0], "a") self.assertEqual(typo[-1], "c") class TestIntroduceTextErrors(unittest.TestCase): """Tests for the introduce_text_errors augmentation function.""" SAMPLE_HTML = """Test

This paragraph contains several important words that should be eligible for modification.

Another sentence with different vocabulary and longer words throughout the document.

""" def test_returns_requested_number_of_typos(self): rng = random.Random(42) _, records = introduce_text_errors(self.SAMPLE_HTML, rng, num_errors=3) self.assertEqual(len(records), 3) def test_typos_appear_in_modified_html(self): rng = random.Random(42) modified_html, records = introduce_text_errors(self.SAMPLE_HTML, rng, num_errors=3) for record in records: self.assertIn(record["typo_word"], modified_html) def test_originals_replaced_in_modified_html(self): rng = random.Random(42) modified_html, records = introduce_text_errors(self.SAMPLE_HTML, rng, num_errors=3) for record in records: # The original word at the typo position should no longer match self.assertNotEqual(record["original_word"], record["typo_word"]) def test_record_fields(self): rng = random.Random(42) _, records = introduce_text_errors(self.SAMPLE_HTML, rng, num_errors=1) self.assertEqual(len(records), 1) self.assertIn("original_word", records[0]) self.assertIn("typo_word", records[0]) def test_no_body_returns_unchanged(self): html = "" rng = random.Random(42) modified, records = introduce_text_errors(html, rng, num_errors=3) self.assertEqual(records, []) def test_num_errors_capped_by_candidates(self): # Only one eligible word html = "

Hello tiny text ok.

" rng = random.Random(42) _, records = introduce_text_errors(html, rng, num_errors=10) # "Hello" and "tiny" and "text" are <5 chars; only "Hello" is 5 chars self.assertLessEqual(len(records), 10) def test_zero_errors_returns_unchanged(self): rng = random.Random(42) modified, records = introduce_text_errors(self.SAMPLE_HTML, rng, num_errors=0) self.assertEqual(records, []) def test_deterministic_with_same_seed(self): rng1 = random.Random(99) modified1, records1 = introduce_text_errors(self.SAMPLE_HTML, rng1, num_errors=3) rng2 = random.Random(99) modified2, records2 = introduce_text_errors(self.SAMPLE_HTML, rng2, num_errors=3) self.assertEqual(records1, records2) self.assertEqual(modified1, modified2) class TestSkipAncestors(unittest.TestCase): """Tests that typos skip headers, footers, tables, and other excluded elements.""" def _get_typo_words(self, html, num_errors=10): rng = random.Random(42) _, records = introduce_text_errors(html, rng, num_errors=num_errors) return {r["original_word"] for r in records} def test_skips_table_content(self): html = """
Eligible looking words inside table

These words should definitely be eligible outside.

""" words = self._get_typo_words(html) self.assertNotIn("Eligible", words) self.assertNotIn("inside", words) self.assertNotIn("table", words) self.assertNotIn("looking", words) def test_skips_header_tags(self): html = """

Important Header Title

Another Subtitle Value

These words should definitely be eligible outside.

""" words = self._get_typo_words(html) self.assertNotIn("Important", words) self.assertNotIn("Header", words) self.assertNotIn("Title", words) self.assertNotIn("Another", words) self.assertNotIn("Subtitle", words) self.assertNotIn("Value", words) def test_skips_footer_element(self): html = """

Paragraph words should definitely qualify here.

""" words = self._get_typo_words(html) self.assertNotIn("Footer", words) self.assertNotIn("content", words) self.assertNotIn("never", words) self.assertNotIn("appear", words) self.assertNotIn("inside", words) def test_skips_header_element(self): html = """
Header content should never appear.

Paragraph words should definitely appear here.

""" words = self._get_typo_words(html) self.assertNotIn("Header", words) def test_skips_script_and_style(self): html = """

Paragraph words should definitely appear here.

""" words = self._get_typo_words(html) self.assertNotIn("doSomething", words) self.assertNotIn("function", words) self.assertNotIn("something", words) def test_skips_sup_and_sub(self): html = """

Normal wordssuperscript and subscript content.

""" words = self._get_typo_words(html) self.assertNotIn("superscript", words) self.assertNotIn("subscript", words) def test_skips_code_and_pre(self): html = """ inlineCode 
preformatted block content

Paragraph words should definitely appear here.

""" words = self._get_typo_words(html) self.assertNotIn("inlineCode", words) self.assertNotIn("preformatted", words) self.assertNotIn("block", words) def test_skips_page_header_class(self): html = """
Header content skipped

Paragraph words should definitely appear here.

""" words = self._get_typo_words(html) self.assertNotIn("Header", words) self.assertNotIn("skipped", words) def test_skips_page_footer_class(self): html = """
Footer content skipped

Paragraph words should definitely appear here.

""" words = self._get_typo_words(html) self.assertNotIn("Footer", words) self.assertNotIn("skipped", words) def test_skips_page_number_class(self): html = """ PageNumber

Paragraph words should definitely appear here.

""" words = self._get_typo_words(html) self.assertNotIn("PageNumber", words) def test_eligible_paragraph_text_is_found(self): html = """

These words should definitely appear as eligible candidates.

""" words = self._get_typo_words(html) # At least some 5+ char words from the paragraph should be candidates self.assertTrue(len(words) > 0) class TestWordFiltering(unittest.TestCase): """Tests that only appropriate words are selected as candidates.""" def test_short_words_skipped(self): html = """

The cat sat on a mat but extraordinary.

""" rng = random.Random(42) _, records = introduce_text_errors(html, rng, num_errors=10) originals = {r["original_word"] for r in records} # Words < 5 chars should not be typo'd for word in ["The", "cat", "sat", "on", "a", "mat", "but"]: self.assertNotIn(word, originals) def test_non_ascii_words_skipped(self): html = """

Héllo wörld but eligible words remain here.

""" rng = random.Random(42) _, records = introduce_text_errors(html, rng, num_errors=10) originals = {r["original_word"] for r in records} # Non-ascii words like Héllo, wörld won't match [A-Za-z]+ fully self.assertNotIn("Héllo", originals) self.assertNotIn("wörld", originals) def test_only_alpha_words(self): html = """

word123 and numbers456 but eligible words remain here.

""" rng = random.Random(42) _, records = introduce_text_errors(html, rng, num_errors=10) originals = {r["original_word"] for r in records} # Regex [A-Za-z]+ splits on digits, so "word" (4 chars) and "numbers" are separate self.assertNotIn("word123", originals) self.assertNotIn("numbers456", originals) class TestTypoTestValidation(unittest.TestCase): """Test that typo records can be validated with TextPresenceTest logic.""" def test_typo_present_in_augmented_absent_in_original(self): from olmocr.bench.tests import TestType, TextPresenceTest html = """

This paragraph contains several important words that should be eligible for modification throughout.

Another sentence with different vocabulary and longer words throughout the document entirely.

""" rng = random.Random(42) original_html = html modified_html, records = introduce_text_errors(html, rng, num_errors=3) self.assertGreater(len(records), 0) # Use the raw HTML text as a stand-in for markdown (avoids importing mine_html_templates) original_text = original_html augmented_text = modified_html validated = 0 for record in records: typo_word = record["typo_word"] test_obj = TextPresenceTest( pdf="test.pdf", page=1, id="test", type=TestType.PRESENT.value, text=typo_word, max_diffs=0, ) passed_original, _ = test_obj.run(original_text) passed_augmented, _ = test_obj.run(augmented_text) if not passed_original and passed_augmented: validated += 1 # At least some typos should pass validation self.assertGreater(validated, 0) class TestTypoDeduplication(unittest.TestCase): """Test that existing presence tests for typo words get deduplicated.""" def test_existing_presence_tests_for_typo_words_removed(self): from olmocr.bench.tests import TestType typo_records = [ {"original_word": "important", "typo_word": "improtant"}, {"original_word": "eligible", "typo_word": "eligble"}, ] typo_words = {r["typo_word"] for r in typo_records} # Simulate tests list that generate_tests_from_html might produce, # including a presence test that matches a typo word tests = [ {"type": TestType.PRESENT.value, "text": "improtant", "id": "dup1"}, {"type": TestType.PRESENT.value, "text": "something_else", "id": "keep1"}, {"type": TestType.TABLE.value, "text": "improtant", "id": "keep2"}, {"type": TestType.PRESENT.value, "text": "eligble", "id": "dup2"}, ] # Apply the same dedup logic used in mine_html_templates filtered = [t for t in tests if not (t.get("type") == TestType.PRESENT.value and t.get("text") in typo_words)] self.assertEqual(len(filtered), 2) ids = {t["id"] for t in filtered} self.assertIn("keep1", ids) self.assertIn("keep2", ids) self.assertNotIn("dup1", ids) self.assertNotIn("dup2", ids) def test_non_present_tests_unaffected(self): from olmocr.bench.tests import TestType typo_words = {"improtant"} tests = [ {"type": TestType.TABLE.value, "text": "improtant", "id": "t1"}, {"type": TestType.ORDER.value, "text": "improtant", "id": "t2"}, ] filtered = [t for t in tests if not (t.get("type") == TestType.PRESENT.value and t.get("text") in typo_words)] self.assertEqual(len(filtered), 2) if __name__ == "__main__": unittest.main()