层3 E1: similarity.py——语义相似度 φ(rouge1 多重集/edit_distance 词级)+ 式(3) k_sem 聚合,21 单测

对应 docs/04 §4 E1。三处对参考实现的替代:edit 吃 str 内部按词切(不再比
token id,回归 tokenizer 无关)、rouge1 集合改多重集(ROUGE-1 标准定义,
数学文本重复词多)、去掉 1e-8 分母平滑(全同串精确得 1)。φ 默认
edit_distance 对齐论文 §5.1。

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

tests/test_similarity.py

@@ -0,0 +1,142 @@
+"""similarity.py 单测——docs/04 §5.1:手构字符串钉死 φ 与 k_sem(式3)。
+
+全部本地 CPU、纯标准库,不依赖 torch/transformers。
+关键手算用例在各测试的注释里逐步展开,方便对着验算。
+"""
+
+import math
+
+import pytest
+
+from ars_opd.similarity import aggregate_similarity, edit_similarity, phi, rouge1
+
+# ---------------------------------------------------------------- rouge1
+
+
+def test_rouge1_identical_is_exact_one():
+    # 全同串必须精确 = 1.0(参考实现因分母 +1e-8 只能得 ≈0.99999998)
+    s = "so x = 5 and y = 12"
+    assert rouge1(s, s) == 1.0
+
+
+def test_rouge1_disjoint_is_zero():
+    assert rouge1("a b c", "x y z") == 0.0
+
+
+def test_rouge1_partial_overlap_hand_computed():
+    # hyp = {a, b, c}, ref = {a, b, d}:overlap = 2
+    # precision = 2/3, recall = 2/3, F1 = 2·(2/3)(2/3) / (4/3) = 2/3
+    assert math.isclose(rouge1("a b c", "a b d"), 2 / 3)
+
+
+def test_rouge1_multiset_counts_repeats():
+    # 多重集语义:hyp = [x,x,x,x], ref = [x] → overlap = min(4,1) = 1
+    # precision = 1/4, recall = 1/1, F1 = 2·(1/4)/(5/4) = 0.4
+    # (参考实现的 set 版会给满分 1.0——数学文本重复词多,这是关键失真点)
+    assert math.isclose(rouge1("x x x x", "x"), 0.4)
+
+
+def test_rouge1_is_bag_of_words_order_blind():
+    # 词袋:只看用了哪些词,不看顺序
+    assert rouge1("a b", "b a") == 1.0
+
+
+def test_rouge1_empty_sides():
+    assert rouge1("", "a b") == 0.0
+    assert rouge1("a b", "") == 0.0
+    assert rouge1("", "") == 0.0
+    assert rouge1("   ", "a") == 0.0  # 纯空白 split 后无词
+
+
+# ---------------------------------------------------------- edit_similarity
+
+
+def test_edit_identical_is_one():
+    s = "so x = 5 and y = 12"
+    assert edit_similarity(s, s) == 1.0
+
+
+def test_edit_totally_different_is_zero():
+    # ["a","b"] vs ["c","d"]:2 次替换,dist=2, max(m,n)=2 → 1 − 1 = 0
+    assert edit_similarity("a b", "c d") == 0.0
+
+
+def test_edit_single_substitution_hand_computed():
+    # ["a","b","c"] vs ["a","x","c"]:1 次替换,dist=1, max=3 → 2/3
+    assert math.isclose(edit_similarity("a b c", "a x c"), 2 / 3)
+
+
+def test_edit_insertion_hand_computed():
+    # ["a","b"] vs ["a","x","b"]:1 次插入,dist=1, max=3 → 2/3
+    assert math.isclose(edit_similarity("a b", "a x b"), 2 / 3)
+
+
+def test_edit_is_order_sensitive():
+    # ["a","b"] vs ["b","a"]:两次替换 dist=2 → 0.0;与 rouge1 的 1.0 互补
+    assert edit_similarity("a b", "b a") == 0.0
+    assert rouge1("a b", "b a") == 1.0
+
+
+def test_edit_empty_sides():
+    assert edit_similarity("", "") == 1.0  # 零距离
+    assert edit_similarity("a b", "") == 0.0  # 全删
+    assert edit_similarity("", "a b") == 0.0  # 全插
+
+
+def test_edit_asymmetric_lengths():
+    # ["a"] vs ["a","b","c","d"]:3 次插入,dist=3, max=4 → 1/4
+    assert math.isclose(edit_similarity("a", "a b c d"), 1 / 4)
+
+
+# --------------------------------------------------------------------- phi
+
+
+def test_phi_default_is_edit_distance():
+    # 论文 §5.1 默认;"a b" vs "b a" 恰能区分两度量(edit=0, rouge1=1)
+    assert phi("a b", "b a") == edit_similarity("a b", "b a") == 0.0
+
+
+def test_phi_dispatch():
+    h, r = "a b c", "a b d"
+    assert phi(h, r, metric="rouge1") == rouge1(h, r)
+    assert phi(h, r, metric="edit_distance") == edit_similarity(h, r)
+
+
+def test_phi_unknown_metric_raises():
+    with pytest.raises(ValueError, match="bleu"):
+        phi("a", "a", metric="bleu")
+
+
+# ----------------------------------------------------- aggregate_similarity
+
+
+def test_aggregate_is_sum_of_phi():
+    # 式(3) 手算:rollouts 与 "a b c" 的 edit 相似度分别为 1.0, 2/3, 0.0
+    chunk = "a b c"
+    rollouts = ["a b c", "a x c", "x y z"]
+    expected = 1.0 + 2 / 3 + 0.0
+    assert math.isclose(aggregate_similarity(chunk, rollouts), expected)
+
+
+def test_aggregate_bounds():
+    # k_sem ∈ [0, N]:全同 → N,全不同 → 0
+    n = 5
+    assert aggregate_similarity("a b", ["a b"] * n) == float(n)
+    assert aggregate_similarity("a b", ["x y"] * n) == 0.0
+
+
+def test_aggregate_is_continuous_soft_count():
+    # φ 连续 ⇒ k_sem 非整数是常态(区别于 token 精确匹配的硬计数)
+    k = aggregate_similarity("a b c", ["a b c", "a x c"])
+    assert 1.0 < k < 2.0
+
+
+def test_aggregate_empty_rollouts():
+    assert aggregate_similarity("a b", []) == 0.0
+
+
+def test_aggregate_metric_passthrough():
+    # "a b" vs "b a":edit 全零,rouge1 全满——验证 metric 真的传下去了
+    chunk, rollouts = "a b", ["b a", "b a"]
+    assert aggregate_similarity(chunk, rollouts, metric="edit_distance") == 0.0
+    assert aggregate_similarity(chunk, rollouts, metric="rouge1") == 2.0