Rebuild from a fresh LeetCode crawl: cleaned tests, exact constraints, public_tests and hints

README.md

@@ -8,25 +8,35 @@ task_categories:
 
 # LeetCode multilingual benchmark dataset
 
-Post-interface rows for the msl-multilingual-self-learning benchmark, flattened to
-one row per (problem, language). Each row has the single `interface` for its
-`language` (python, cpp, go, java, rust, javascript, typescript, php, ruby), plus
-the shared `canonical_tests` (a Python check(candidate) oracle), `problem_description`,
-and `metadata`. Stored as `data/<split>/<lang>-NNN.jsonl` (row-boundary shards so
-every file stays under HF's ~10 MiB per-file ceiling); HF globs them into a single
-split.
+LeetCode problems for the msl-multilingual-self-learning benchmark, flattened to
+one row per (problem, language) in 9 languages, stored as `data/<split>/<lang>-NNN.jsonl`.
+Each row has the `interface` for its `language` (from LeetCode's code snippets),
+the shared `canonical_tests` (Python asserts from newfacade/LeetCodeDataset),
+the `problem_description` (from the LeetCode page) and `metadata`.
+
+Tests that break the problem's Constraints, do not fit a declared type in some
+language, or expect inf/nan were removed; problems that take trees or linked
+lists, modify their input in place, or kept fewer than 10 tests were dropped.
+The test split also drops problems with several valid answers or whose text
+refers to a figure. Where answers may come in any order or are decimals,
+the asserts call `answers_match` (defined at the top of the test source), and
+`canonical_tests.comparison` says which rule applies ("unordered", "float" or "exact").
+`public_tests` holds the inputs of LeetCode's Testcase panel (one string per
+case, one JSON value per line in parameter order; LeetCode publishes no
+outputs), and `hints` the page's hints as plain text.
+`reports/dropped.md` lists every dropped problem and test count, and
+`reports/<split>.json` has the details.
 
 ## Splits
 
-- `train`: 2274 problems x 9 languages (2641 in source; 367 lack complete
-  interfaces).
-- `test`: 201 problems x 9 languages = 1809 rows (held-out eval).
+- `train`: 2103 problems x 9 languages = 18927 rows (2641 in the source dataset)
+- `test`: 191 problems x 9 languages = 1719 rows (228 in the source dataset)
 
 ## Load
 
 ```python
 from datasets import load_dataset
-train = load_dataset("neulab/leetcode", split="train")  # one Dataset, all rows
-test  = load_dataset("neulab/leetcode", split="test")   # one Dataset, 1809 rows
+train = load_dataset("neulab/leetcode", split="train")
+test  = load_dataset("neulab/leetcode", split="test")
 python_rows = test.filter(lambda r: r["language"] == "python")
 ```