parse-bench 1.0.2__tar.gz → 1.0.3__tar.gz

{parse_bench-1.0.2 → parse_bench-1.0.3}/CHANGELOG.md

@@ -4,6 +4,14 @@ All notable changes to `parse-bench` are recorded here. The format follows
 [Keep a Changelog](https://keepachangelog.com/en/1.1.0/) and the project uses
 [Semantic Versioning](https://semver.org/).
 
+## [Unreleased]
+
+### Changed
+- Dropped the `scikit-learn` dependency. Layout-detection average precision is
+  now computed with a small numpy implementation that matches
+  `sklearn.metrics.average_precision_score` (including tied-score handling),
+  so evaluation numbers are unchanged. `scipy` remains a core dependency.
+
 ## [1.0.2] - 2026-09-04
 
 Ports two fixes that landed in the internal harness while 1.0.1 was being cut.

{parse_bench-1.0.2 → parse_bench-1.0.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: parse-bench
-Version: 1.0.2
+Version: 1.0.3
 Summary: ParseBench: a benchmark and evaluation harness for document parsing systems on real-world enterprise documents
 Project-URL: Homepage, https://parsebench.ai
 Project-URL: Repository, https://github.com/run-llama/ParseBench
@@ -43,7 +43,6 @@ Requires-Dist: python-dotenv>=1.0.0
 Requires-Dist: python-levenshtein>=0.25.0
 Requires-Dist: rapidfuzz>=3.0.0
 Requires-Dist: rich>=13.7.0
-Requires-Dist: scikit-learn>=1.8.0
 Requires-Dist: scipy>=1.16.3
 Requires-Dist: unidecode>=1.3.0
 Provides-Extra: anthropic

{parse_bench-1.0.2 → parse_bench-1.0.3}/pyproject.toml

@@ -41,7 +41,6 @@ dependencies = [
     "python-Levenshtein>=0.25.0",
     "rapidfuzz>=3.0.0",
     "rich>=13.7.0",
-    "scikit-learn>=1.8.0",
     "scipy>=1.16.3",
     "unidecode>=1.3.0",
     "anls-star>=0.1.0",
@@ -199,7 +198,6 @@ module = [
     "pandas",
     "fuzzysearch",
     "scipy.*",
-    "sklearn.*",
     "lxml.*",
     "datalab_sdk",
     "datalab_sdk.*",

{parse_bench-1.0.2 → parse_bench-1.0.3}/src/parse_bench/__init__.py

@@ -1,3 +1,3 @@
 """Document parsing evaluation system for competitive benchmarking."""
 
-__version__ = "1.0.2"
+__version__ = "1.0.3"

{parse_bench-1.0.2 → parse_bench-1.0.3}/src/parse_bench/evaluation/metrics/layoutdet/classification_utils.py

@@ -8,13 +8,44 @@ from collections import defaultdict
 from collections.abc import Callable
 
 import numpy as np
-from sklearn.metrics import average_precision_score
 
 from parse_bench.evaluation.metrics.layoutdet.iou import compute_iou_matrix
 
 OverlapFn = Callable[[dict, dict], float]
 
 
+def average_precision_score(y_true: np.ndarray, y_score: np.ndarray) -> float:
+    """
+    Binary average precision, matching ``sklearn.metrics.average_precision_score``.
+
+    AP = sum_n (R_n - R_{n-1}) * P_n, where precision/recall are evaluated at each
+    distinct score threshold (tied scores are grouped, as in scikit-learn).
+
+    :param y_true: Binary labels, shape (N,)
+    :param y_score: Confidence scores, shape (N,)
+    :return: Average precision in [0, 1]; 0.0 if there are no positives
+    """
+    y_true = np.asarray(y_true, dtype=float).ravel()
+    y_score = np.asarray(y_score, dtype=float).ravel()
+    n_pos = float(y_true.sum())
+    if len(y_true) == 0 or n_pos == 0:
+        return 0.0
+
+    order = np.argsort(-y_score, kind="stable")
+    y_true = y_true[order]
+    y_score = y_score[order]
+
+    # Indices of the last element in each group of tied scores
+    distinct = np.where(np.diff(y_score) != 0)[0]
+    threshold_idx = np.concatenate([distinct, [len(y_true) - 1]])
+
+    tps = np.cumsum(y_true)[threshold_idx]
+    precision = tps / (threshold_idx + 1)
+    recall = tps / n_pos
+    prev_recall = np.concatenate([[0.0], recall[:-1]])
+    return float(np.sum((recall - prev_recall) * precision))
+
+
 def match_predictions_to_gt(
     pred_boxes: np.ndarray,  # (N, 4) xyxy format
     pred_scores: np.ndarray,  # (N,)

parse_bench-1.0.3/tests/parse_bench/evaluation/metrics/layoutdet/test_average_precision.py

@@ -0,0 +1,30 @@
+"""Pins the numpy average-precision implementation to scikit-learn's semantics."""
+
+import numpy as np
+import pytest
+
+from parse_bench.evaluation.metrics.layoutdet.classification_utils import average_precision_score
+
+
+@pytest.mark.parametrize(
+    ("y_true", "y_score", "expected"),
+    [
+        # Reference values computed with sklearn.metrics.average_precision_score
+        ([1, 0, 1, 1], [0.9, 0.8, 0.7, 0.6], 0.8055555555555556),
+        ([0, 0, 1, 1], [0.1, 0.4, 0.35, 0.8], 0.8333333333333333),
+        ([1, 1, 0, 0], [0.9, 0.8, 0.7, 0.6], 1.0),
+        ([0, 0, 1, 1], [0.9, 0.8, 0.7, 0.6], 0.41666666666666663),
+        # Tied scores are grouped at one threshold (sklearn semantics)
+        ([1, 0, 1, 0], [0.5, 0.5, 0.5, 0.5], 0.5),
+        ([1, 0, 0, 1], [0.9, 0.5, 0.5, 0.5], 0.75),
+    ],
+)
+def test_matches_sklearn_reference_values(y_true, y_score, expected):
+    ap = average_precision_score(np.array(y_true, dtype=float), np.array(y_score, dtype=float))
+    assert ap == pytest.approx(expected)
+
+
+def test_degenerate_inputs():
+    assert average_precision_score(np.array([]), np.array([])) == 0.0
+    assert average_precision_score(np.zeros(3), np.array([0.3, 0.2, 0.1])) == 0.0
+    assert average_precision_score(np.ones(3), np.array([0.3, 0.2, 0.1])) == 1.0