@furongjun1999/dsh-memory 0.9.6 → 0.9.8
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +150 -662
- package/codebuddy/CODEBUDDY.md +333 -204
- package/data/policy.json +38 -38
- package/docs/README.md +18 -0
- package/docs/discipline/harnesses.yaml +11 -3
- package/docs/discipline/templates/compact.txt.tmpl +20 -18
- package/docs/discipline/templates/full.md.tmpl +63 -61
- package/docs/discipline/templates/rules.mdc.tmpl +5 -3
- package/docs/discipline/templates/skill.md.tmpl +25 -23
- package/docs/discipline/templates/spec.txt.tmpl +2 -2
- package/docs/discipline/templates/zcode-user.md.tmpl +1 -1
- package/dsh/cordis.yml.example +139 -139
- package/lib/hooks.d.ts +19 -0
- package/lib/hooks.js +30 -7
- package/lib/index.js +18 -3
- package/lib/init.js +19 -4
- package/lib/lib/hook_audit.js +64 -2
- package/lib/lib/prompt_safety.js +14 -3
- package/lib/tools.d.ts +7 -0
- package/lib/tools.js +22 -1
- package/md_cg/_guard_kit.py +121 -121
- package/md_cg/anchor.py +114 -114
- package/md_cg/audit.py +15 -2
- package/md_cg/chain.py +3 -2
- package/md_cg/config_registry_guard.py +487 -487
- package/md_cg/conformance.py +747 -747
- package/md_cg/consolidate.py +1793 -1793
- package/md_cg/crypto.py +794 -794
- package/md_cg/hyperedge.py +262 -262
- package/md_cg/interface_schema.py +146 -146
- package/md_cg/lifecycle.py +426 -426
- package/md_cg/mdcos.py +12 -8
- package/md_cg/predict.py +1103 -1103
- package/md_cg/protect.py +654 -654
- package/md_cg/provenance.py +606 -606
- package/md_cg/queue_health.py +130 -130
- package/md_cg/refine.py +613 -613
- package/md_cg/retrieve.py +128 -128
- package/md_cg/revise.py +140 -140
- package/md_cg/rollback.py +657 -657
- package/md_cg/rollback_cli.py +124 -124
- package/md_cg/rotate.py +346 -346
- package/md_cg/routing.py +490 -490
- package/md_cg/self_surface.py +50 -50
- package/md_cg/semantic/canonical.py +144 -144
- package/md_cg/semantic/zh_en_atoms.py +141 -141
- package/md_cg/sources.py +1091 -1091
- package/md_cg/stg.py +3 -2
- package/md_cg/sustain.py +10 -1
- package/md_cg/test_audit_rotate_parity.py +279 -279
- package/md_cg/test_autonomy_modes.py +2013 -2013
- package/md_cg/test_cn_recall_grams.py +206 -206
- package/md_cg/test_code_kind_forbidden.py +101 -0
- package/md_cg/test_condaccept.py +156 -156
- package/md_cg/test_condgate.py +141 -141
- package/md_cg/test_heartbeat_ledger.py +9 -3
- package/md_cg/test_i50e_readside_protection.py +658 -658
- package/md_cg/test_issue111_gate_order.py +76 -76
- package/md_cg/test_issue111_scoring_contract.py +86 -86
- package/md_cg/test_issue121_calibration_bounds.py +293 -293
- package/md_cg/test_issue145_sleep_no_silent_rollback.py +167 -167
- package/md_cg/test_issue147_pack_id_escape.py +378 -378
- package/md_cg/test_issue148_merge_downgrade_leak.py +192 -192
- package/md_cg/test_issue149_accel_default_face.py +248 -248
- package/md_cg/test_issue150_empty_query_terms.py +328 -328
- package/md_cg/test_issue151_153_157_write_cache_index.py +151 -151
- package/md_cg/test_issue152_sealed_postings.py +309 -309
- package/md_cg/test_issue153b_index_generation_invalidation.py +142 -142
- package/md_cg/test_issue157b_index_remainder.py +344 -344
- package/md_cg/test_issue189_conclusion_bidirectional.py +223 -223
- package/md_cg/test_issue191_import_trust.py +382 -382
- package/md_cg/test_issue193_write_scope.py +378 -378
- package/md_cg/test_issue212_writeface_gaps.py +349 -349
- package/md_cg/test_issue214_reject_sensitivity.py +266 -266
- package/md_cg/test_issue215_condition_source_gap.py +272 -272
- package/md_cg/test_issue216_clash_routes.py +214 -214
- package/md_cg/test_issue217_slot_routes.py +248 -248
- package/md_cg/test_issue218_postcondition_value_blind.py +222 -222
- package/md_cg/test_issue258_pack_sensitivity.py +468 -468
- package/md_cg/test_issue265_coldverify_multiproc.py +438 -438
- package/md_cg/test_issue274_read_node_id_none.py +341 -341
- package/md_cg/test_issue275_layer_param.py +250 -250
- package/md_cg/test_issue277_node_id_traversal.py +384 -384
- package/md_cg/test_issue278_neg_separated_tension.py +271 -271
- package/md_cg/test_issue278_neg_window.py +294 -294
- package/md_cg/test_issue278_negative_layering.py +83 -83
- package/md_cg/test_issue284_skip_not_pass.py +291 -291
- package/md_cg/test_issue294_flywheel_sensitivity.py +573 -573
- package/md_cg/test_issue301_baseline_corpus_gate.py +245 -245
- package/md_cg/test_issue303_design_boundaries.py +262 -262
- package/md_cg/test_issue315_merge_branch_sensitivity.py +383 -383
- package/md_cg/test_issue319_tenant_unregistered_bypass.py +256 -256
- package/md_cg/test_issue321_mcp_deep_json.py +304 -304
- package/md_cg/test_issue322_node_write_lock.py +727 -727
- package/md_cg/test_issue324_write_size_fields.py +214 -214
- package/md_cg/test_issue325_audit_poison.py +229 -229
- package/md_cg/test_issue326_ingest_auto_whitelist.py +331 -331
- package/md_cg/test_issue326_key_perms.py +355 -355
- package/md_cg/test_issue326_twin_parser_parity.py +323 -323
- package/md_cg/test_issue326_vm_call_frame.py +569 -569
- package/md_cg/test_issue345_writelimit_effective_sens.py +233 -233
- package/md_cg/test_issue346_overwrite_sensitivity_gate.py +424 -424
- package/md_cg/test_issue348_insight_verify_op.py +199 -199
- package/md_cg/test_issue348_merge_rmw_lock.py +345 -345
- package/md_cg/test_issue350_casefold_collision.py +235 -235
- package/md_cg/test_issue351_ingest_root_fallback.py +201 -201
- package/md_cg/test_issue352_imgaudit.py +418 -418
- package/md_cg/test_issue352_pathsjson.py +537 -537
- package/md_cg/test_issue352_refindex.py +412 -412
- package/md_cg/test_issue352_rollback.py +604 -604
- package/md_cg/test_issue354_p2_writelimit_toctou.py +327 -327
- package/md_cg/test_issue354_rust_index_containment.py +571 -571
- package/md_cg/test_issue354_rust_wal_containment.py +606 -606
- package/md_cg/test_issue358_truncation_observable.py +247 -247
- package/md_cg/test_issue360_tidy_private_excerpt.py +195 -195
- package/md_cg/test_issue361_blindspot_not_silent_drop.py +494 -494
- package/md_cg/test_issue362_neg_condition.py +468 -468
- package/md_cg/test_issue363_lexical_idf.py +275 -275
- package/md_cg/test_issue370_readcache_assembly.py +488 -488
- package/md_cg/test_issue376_heartbeat_failed.py +293 -293
- package/md_cg/test_issue377_layer_protect_stamp.py +324 -324
- package/md_cg/test_issue378_protected_floor.py +310 -310
- package/md_cg/test_issue383_state_event_write_gate.py +462 -462
- package/md_cg/test_issue397_sens_eff_wiring.py +435 -435
- package/md_cg/test_issue398_toolface_op_parity.py +511 -511
- package/md_cg/test_issue399_hive_json_shape.py +317 -317
- package/md_cg/test_issue401_wisdom_codegen.py +295 -295
- package/md_cg/test_issue404_heal_receipt.py +389 -389
- package/md_cg/test_issue409_compact_index_completeness.py +509 -509
- package/md_cg/test_issue410_verify_demote_write.py +366 -366
- package/md_cg/test_issue411_scrub_idempotence.py +575 -575
- package/md_cg/test_issue412_discipline_producer_parity.py +311 -311
- package/md_cg/test_issue413_preimage_failclosed.py +239 -239
- package/md_cg/test_issue413_preimage_twin.py +339 -339
- package/md_cg/test_issue415_deps_gate_probe.py +357 -357
- package/md_cg/test_issue416_entry_key_subset.py +525 -525
- package/md_cg/test_issue417_twophase_merge_landed.py +423 -423
- package/md_cg/test_issue425_ingest_library_path.py +331 -331
- package/md_cg/test_issue426c_enqueue_not_silent.py +346 -346
- package/md_cg/test_issue426e_restore_overwrite_live.py +310 -310
- package/md_cg/test_issue426f_paths_gbk_failclosed.py +301 -301
- package/md_cg/test_issue427_serve_status_shape.py +191 -191
- package/md_cg/test_issue431_432_468_protect_single_point.py +322 -322
- package/md_cg/test_issue433_outlook_threshold_single_point.py +278 -278
- package/md_cg/test_issue434_435_float_gate.py +90 -90
- package/md_cg/test_issue439_sibling_sensitivity_inherit.py +448 -448
- package/md_cg/test_issue43_default_policy.py +874 -874
- package/md_cg/test_issue444_scrub_evidence_plaintext.py +281 -281
- package/md_cg/test_issue445_metacognition_ledger_gate.py +454 -454
- package/md_cg/test_issue446_453_403_verified_evidence.py +382 -382
- package/md_cg/test_issue450_sensitivity_fallback_reason.py +51 -51
- package/md_cg/test_issue450_writegate_unknown_sens.py +203 -203
- package/md_cg/test_issue451_audit_actor.py +313 -313
- package/md_cg/test_issue452_ingest_root_sensitivity.py +299 -299
- package/md_cg/test_issue463_ghost_prune_guard.py +259 -259
- package/md_cg/test_issue466_basis_registry.py +469 -469
- package/md_cg/test_issue467_identity_sensitivity_parity.py +360 -360
- package/md_cg/test_issue469_scrub_override_passthrough.py +126 -126
- package/md_cg/test_issue470_wal_event_durable.py +427 -427
- package/md_cg/test_issue471_identity_readable.py +136 -136
- package/md_cg/test_issue472_audit_chain.py +90 -90
- package/md_cg/test_issue472_prune_ledger.py +105 -105
- package/md_cg/test_issue472_rollback_degraded.py +81 -81
- package/md_cg/test_issue491_compose_sandbox.py +460 -460
- package/md_cg/test_issue498_merge_trace_stale_index.py +280 -280
- package/md_cg/test_issue511_scoreboard_bad_files.py +256 -256
- package/md_cg/test_issue521_rotate_lock_suffix.py +110 -110
- package/md_cg/test_issue524_orphan_noise_no_demote.py +88 -88
- package/md_cg/test_issue525_read_null_semantics.py +52 -52
- package/md_cg/test_issue547_append_jsonl_whitelist.py +114 -114
- package/md_cg/test_issue547_audit_durable.py +96 -96
- package/md_cg/test_issue557_r2_write_entries_size_gate.py +83 -83
- package/md_cg/test_issue559_refine_prefix.py +69 -69
- package/md_cg/test_issue562_retired_faces.py +150 -0
- package/md_cg/test_issue71_queue_backlog.py +376 -376
- package/md_cg/test_issue72_coldverify_controlled_drain.py +420 -420
- package/md_cg/test_issue82_credential_rules.py +271 -271
- package/md_cg/test_issue91_verifier_sandbox.py +314 -314
- package/md_cg/test_issue93_remember_redaction.py +282 -282
- package/md_cg/test_logref.py +1141 -1141
- package/md_cg/test_md_access_parity.py +330 -330
- package/md_cg/test_md_cg_no_hive_dep.py +139 -139
- package/md_cg/test_mutation_residue.py +119 -119
- package/md_cg/test_mutation_rollback.py +835 -835
- package/md_cg/test_n123_derive_expiry_chain.py +210 -210
- package/md_cg/test_n184_keys_concurrent_provision.py +1 -1
- package/md_cg/test_n209_verify_write_face_gates.py +466 -466
- package/md_cg/test_n62_tenant_bind_failclosed.py +200 -200
- package/md_cg/test_neg_tail_honesty.py +733 -733
- package/md_cg/test_p21_tokens.py +294 -294
- package/md_cg/test_p27_docindex.py +792 -792
- package/md_cg/test_p28_refcheck.py +1053 -1053
- package/md_cg/test_p2_six_elements.py +476 -476
- package/md_cg/test_p33_ccg_wiring.py +300 -300
- package/md_cg/test_p3_legacy_closure.py +438 -438
- package/md_cg/test_p40_refine_worklist.py +244 -244
- package/md_cg/test_p4_freshness.py +781 -781
- package/md_cg/test_p9c_dedup_hints.py +288 -288
- package/md_cg/test_policy_required_ccg.py +444 -444
- package/md_cg/test_read_scope_b27.py +2 -2
- package/md_cg/test_retr_s5.py +312 -312
- package/md_cg/test_s1_ast_judge.py +109 -109
- package/md_cg/test_security_audit_v21.py +249 -249
- package/md_cg/test_state_event_op.py +521 -521
- package/md_cg/test_stgidx_index_parity.py +1024 -1024
- package/md_cg/test_sustain_bad_stamp.py +178 -0
- package/md_cg/test_sustain_bounded.py +652 -652
- package/md_cg/test_trust_edge_reachability.py +405 -405
- package/md_cg/test_w7_redlines.py +659 -659
- package/md_cg/test_write_return_shape.py +919 -919
- package/md_cg/test_writepipe.py +228 -228
- package/md_cg/trust.py +1114 -1114
- package/md_cg/twophase.py +303 -303
- package/md_cg/whitebox.py +536 -536
- package/md_cg/whitebox_kb/wisdom/compose_engine.py +1634 -1634
- package/md_cg/whitebox_kb/wisdom/core/condition_vm.py +318 -318
- package/md_cg/whitebox_kb/wisdom/core/parser.py +1082 -1082
- package/md_cg/whitebox_kb/wisdom/md_access.py +490 -490
- package/md_cg/whitebox_kb/wisdom/migrate_wisdom.py +497 -497
- package/md_cg/whitebox_kb/wisdom/mini_python.py +995 -995
- package/md_cg/whitebox_kb/wisdom/net_units.py +2900 -2900
- package/md_cg/whitebox_kb/wisdom/semantic_translate.py +4060 -4060
- package/md_cg/writepipe.py +1337 -1337
- package/package.json +3 -2
- package/scripts/workspace_index.py +328 -0
- package/src/hooks.ts +932 -908
- package/src/index.ts +21 -6
- package/src/init.ts +16 -4
- package/src/lib/hook_audit.ts +48 -2
- package/src/lib/mdcg_client.ts +469 -469
- package/src/lib/prompt_safety.ts +17 -3
- package/src/tools.ts +23 -1
- package/zcode/AGENTS.md +333 -204
|
@@ -1,275 +1,275 @@
|
|
|
1
|
-
# -*- coding: utf-8 -*-
|
|
2
|
-
# 功能名:缺省 lexical 路的真实打分(IDF 加权覆盖度,issue #363)
|
|
3
|
-
# 生效条件:python -X utf8 -m md_cg.test_issue363_lexical_idf [--mutate];被测面=md_cg/mdcg.py 的 idf_of / lexical_sim(df,n_docs) / MdCG._score 的 df 统计与透传
|
|
4
|
-
# 子功能:A 组=缺陷行为面(同分退化 → 稀有 bigram 命中者居首、分数可分辨);K 组=非恒真对照(不传 df 的旧公式逐位不变、jaccard 不受影响、idf 与 bm25 同公式且随 df 单调降、同内容文档不产生虚假分差);E 组=判据单点(idf 算术全仓唯一、df 只在 _score 统计、透传在位)
|
|
5
|
-
# 执行:--mutate 时把修复逐条回退(撤 _score 的 df 透传 / 撤 lexical_sim 的 IDF 分支 / 把 idf 打成常量),守卫必须按指定项转红
|
|
6
|
-
# 验证方式:正向全部 PASS 且退出码 0;--mutate 下红项集合须与预期逐项相符(空集合即不合格),变异锚点未命中 ⇒ 退出码 2(fail-closed)
|
|
7
|
-
# 不适用条件:① 不判 Rust 侧(rust/src/retrieval.rs 的 legacy 公式未同步改,跨侧 rank_parity 由 scripts/rank_parity.py 的 jaccard 口径钉住,本件不覆盖);② 不判 `_score` 之外的二文档比对(S5 负记忆抑制仍走不传 df 的旧公式,K1 只钉其不变);③ 不判跨候选池的分数可比性(df 统计面=本批 docs,跨批分数不可横比,已在 lexical_sim docstring 登记);④ 不判冻结集 hit@k/MRR 的绝对量级(随语料浮动,只登记在报告里,不作断言)。
|
|
8
|
-
"""issue #363 守卫:缺省 lexical 路的真实打分(IDF 加权覆盖度)。
|
|
9
|
-
|
|
10
|
-
病象(修前,本守卫 A 组逐条复现):
|
|
11
|
-
查询「甲甲乙乙」在 40 条只含「甲甲」的干扰 + 1 条只含「乙乙」的稀有文档上,
|
|
12
|
-
legacy 口径 `|q∩d|/|q|` 对 41 条文档全部给 0.333333 —— 分数只有 k/|q| 几个
|
|
13
|
-
离散档,名次由 `(-score, -importance, id)` 的 id 兜底决定=写入序:稀有命中
|
|
14
|
-
的 z_rare 排第 41(探针实测 41/41,分数档数 1/41)。
|
|
15
|
-
|
|
16
|
-
修法(判据单点,不新造第二套分词口径):
|
|
17
|
-
· `idf_of(df, n_docs) = log1p((N-df+0.5)/(df+0.5))`——与 `md_cg/bm25.py`
|
|
18
|
-
的 BM25 IDF 同一算术,全仓唯一实现;
|
|
19
|
-
· `lexical_sim(..., df=, n_docs=)` 在 legacy 口径下走
|
|
20
|
-
`Σ_{g∈q∩d} idf(g) / Σ_{g∈q} idf(g)`;**不传 df 时逐位保持旧公式**;
|
|
21
|
-
· `MdCG._score` 按本批候选文档统计查询 bigram 的 df 并透传(与 bm25 的
|
|
22
|
-
posting 面同为「当前候选池」口径,不扫全库)。
|
|
23
|
-
|
|
24
|
-
判别力来源(非恒真):A1 修前 41 → 修后 1;A2 稀有分/常见分差 > 1e-3;
|
|
25
|
-
K1 旧公式逐位相等(防止「到处都变了」);K4 同内容文档同分(防止「凭空分差」);
|
|
26
|
-
E1/E4 单一实现(防止第二份 IDF 公式漂移)。
|
|
27
|
-
"""
|
|
28
|
-
from __future__ import annotations
|
|
29
|
-
|
|
30
|
-
import math
|
|
31
|
-
import os
|
|
32
|
-
import re
|
|
33
|
-
import sys
|
|
34
|
-
import tempfile
|
|
35
|
-
import types
|
|
36
|
-
|
|
37
|
-
sys.stdout.reconfigure(encoding="utf-8", errors="replace")
|
|
38
|
-
|
|
39
|
-
_HERE = os.path.dirname(os.path.abspath(__file__))
|
|
40
|
-
_REPO = os.path.dirname(_HERE)
|
|
41
|
-
if _REPO not in sys.path:
|
|
42
|
-
sys.path.insert(0, _REPO)
|
|
43
|
-
|
|
44
|
-
_TMP = tempfile.mkdtemp(prefix="iss363_")
|
|
45
|
-
os.environ["MDCG_ROOT"] = os.path.join(_TMP, "root")
|
|
46
|
-
os.environ["MDCG_AUX_ROOT"] = os.path.join(_TMP, "auxroot")
|
|
47
|
-
os.environ["MDCG_MASTER_KEY"] = "cd" * 32
|
|
48
|
-
os.environ["MDCG_TOKEN_FILE"] = os.path.join(_TMP, "tokens.json")
|
|
49
|
-
for _k in ("MDCG_TOKEN", "MDCG_VERIFY_KEY", "MDCG_SCORE_MODE", "MDCG_UNIFY_QUERY",
|
|
50
|
-
"MDCG_CN_GRAMS", "MDCG_READ_CACHE"):
|
|
51
|
-
os.environ.pop(_k, None)
|
|
52
|
-
|
|
53
|
-
from md_cg import mdcg as _mdcg # noqa: E402
|
|
54
|
-
|
|
55
|
-
PASS: list = []
|
|
56
|
-
FAIL: list = []
|
|
57
|
-
|
|
58
|
-
|
|
59
|
-
def ok(cond, msg, detail=""):
|
|
60
|
-
(PASS if cond else FAIL).append(msg)
|
|
61
|
-
print((" PASS " if cond else " FAIL ") + msg
|
|
62
|
-
+ ((" ← " + str(detail)) if (detail and not cond) else ""))
|
|
63
|
-
return bool(cond)
|
|
64
|
-
|
|
65
|
-
|
|
66
|
-
# ---------------------------------------------------------------- 夹具
|
|
67
|
-
|
|
68
|
-
QUERY = "甲甲乙乙"
|
|
69
|
-
COMMON = "甲甲 记录 第%d条"
|
|
70
|
-
RARE = "乙乙 独特 内容"
|
|
71
|
-
|
|
72
|
-
|
|
73
|
-
def _build(cg_cls, root, n_common=40):
|
|
74
|
-
cg = cg_cls(root)
|
|
75
|
-
for i in range(n_common):
|
|
76
|
-
cg.add("a_common%02d" % i, COMMON % i, layer="knowledge")
|
|
77
|
-
cg.add("z_rare", RARE, layer="knowledge")
|
|
78
|
-
cg.flush()
|
|
79
|
-
return cg
|
|
80
|
-
|
|
81
|
-
|
|
82
|
-
def _search_rows(cg, k=41):
|
|
83
|
-
res, meta = cg.search(QUERY, k=k, judge=False, record=False)
|
|
84
|
-
return [(n["id"], float(s)) for n, s, _q in res], meta
|
|
85
|
-
|
|
86
|
-
|
|
87
|
-
def _fixture_readout(cg_cls, root):
|
|
88
|
-
cg = _build(cg_cls, root)
|
|
89
|
-
rows, meta = _search_rows(cg)
|
|
90
|
-
ids = [i for i, _ in rows]
|
|
91
|
-
scores = [s for _, s in rows]
|
|
92
|
-
rank = ids.index("z_rare") + 1 if "z_rare" in ids else 0
|
|
93
|
-
rare = scores[ids.index("z_rare")] if rank else 0.0
|
|
94
|
-
common = max((s for i, s in rows if i.startswith("a_common")), default=0.0)
|
|
95
|
-
return {"rank": rank, "rare": rare, "common": common,
|
|
96
|
-
"levels": len({round(s, 12) for s in scores}), "n": len(rows),
|
|
97
|
-
"meta": meta}
|
|
98
|
-
|
|
99
|
-
|
|
100
|
-
# ---------------------------------------------------------------- A 行为面
|
|
101
|
-
|
|
102
|
-
def group_a():
|
|
103
|
-
print("== A 缺陷行为面:同分退化 → 稀有 bigram 命中者居首 ==")
|
|
104
|
-
r = _fixture_readout(_mdcg.MdCG, os.path.join(_TMP, "a1"))
|
|
105
|
-
print(" 读数:rank=%s/%s rare=%.6f common=%.6f levels=%s"
|
|
106
|
-
% (r["rank"], r["n"], r["rare"], r["common"], r["levels"]))
|
|
107
|
-
ok(r["n"] == 41, "A0 夹具非空转:41 条文档全部参与打分(正对照)", r["n"])
|
|
108
|
-
ok(r["rank"] == 1,
|
|
109
|
-
"A1 稀有 bigram(只出现在 1/41 文档)命中者居首(修前 rank=41,写入序兜底)",
|
|
110
|
-
"rank=%s" % r["rank"])
|
|
111
|
-
ok(r["rare"] - r["common"] > 1e-3,
|
|
112
|
-
"A2 稀有/常见命中分数可分辨(修前逐位相同 0.333333)",
|
|
113
|
-
"rare=%r common=%r" % (r["rare"], r["common"]))
|
|
114
|
-
ok(r["levels"] >= 2, "A3 分数档数 > 1(修前 1/41:整批同分)", r["levels"])
|
|
115
|
-
|
|
116
|
-
|
|
117
|
-
# ---------------------------------------------------------------- K 非恒真对照
|
|
118
|
-
|
|
119
|
-
def group_k():
|
|
120
|
-
print("== K 非恒真对照:旧公式 / jaccard / idf 公式 / 无虚假分差 ==")
|
|
121
|
-
qb = {"甲甲", "乙乙", "甲乙"}
|
|
122
|
-
nb = {"甲甲", "甲乙"}
|
|
123
|
-
# K1 不传 df ⇒ 旧公式逐位不变(S5 等二文档比对路径)
|
|
124
|
-
ok(_mdcg.lexical_sim(qb, nb) == len(qb & nb) / len(qb),
|
|
125
|
-
"K1 不传 df 时 legacy 仍是 |q∩d|/|q|(旧公式逐位不变)",
|
|
126
|
-
_mdcg.lexical_sim(qb, nb))
|
|
127
|
-
ok(_mdcg.lexical_sim(set(), nb) == 0.0 and _mdcg.lexical_sim(qb, set()) == 0.0,
|
|
128
|
-
"K1b 空集合恒 0.0(边界不因 IDF 改变)")
|
|
129
|
-
# K2 jaccard 不受 df 影响
|
|
130
|
-
df = {"甲甲": 40, "乙乙": 1, "甲乙": 40}
|
|
131
|
-
ok(_mdcg.lexical_sim(qb, nb, mode="jaccard", df=df, n_docs=41)
|
|
132
|
-
== len(qb & nb) / len(qb | nb),
|
|
133
|
-
"K2 jaccard 口径不因传入 df 改变(口径隔离)")
|
|
134
|
-
# K3 idf 与 bm25 同公式
|
|
135
|
-
for d in (0, 1, 20, 40, 41):
|
|
136
|
-
want = math.log1p((41 - d + 0.5) / (d + 0.5))
|
|
137
|
-
got = _mdcg.idf_of(d, 41)
|
|
138
|
-
ok(abs(got - want) < 1e-15,
|
|
139
|
-
"K3 idf_of(df=%d, N=41) 与 bm25 公式逐位一致" % d, (got, want))
|
|
140
|
-
ok(_mdcg.idf_of(0, 41) > _mdcg.idf_of(20, 41) > _mdcg.idf_of(40, 41),
|
|
141
|
-
"K3b idf 随 df 单调递减(高频 bigram 权重更低)")
|
|
142
|
-
ok(_mdcg.idf_of(5, 0) == 0.0, "K3c N=0 时 idf 恒 0(不猜测)")
|
|
143
|
-
# K4 同内容文档不得产生虚假分差
|
|
144
|
-
ok(_mdcg.lexical_sim(qb, nb, df=df, n_docs=41)
|
|
145
|
-
== _mdcg.lexical_sim(qb, nb, df=df, n_docs=41),
|
|
146
|
-
"K4 同一输入同分(打分是纯函数)")
|
|
147
|
-
root = os.path.join(_TMP, "k4")
|
|
148
|
-
cg = _mdcg.MdCG(root)
|
|
149
|
-
for i in range(6):
|
|
150
|
-
cg.add("dup%d" % i, "丙丙 同样内容", layer="knowledge")
|
|
151
|
-
cg.flush()
|
|
152
|
-
# 用与内容同词的查询(有 LIKE 命中 ⇒ 走真实打分路径,而不是 0 分兜底池)
|
|
153
|
-
rows = [(n["id"], float(s)) for n, s, _q in
|
|
154
|
-
cg.search("丙丙", k=6, judge=False, record=False)[0]]
|
|
155
|
-
ok(len(rows) == 6 and len({round(s, 12) for _, s in rows}) == 1,
|
|
156
|
-
"K4b 同内容文档分数逐位相同(IDF 不制造虚假分差)",
|
|
157
|
-
[s for _, s in rows])
|
|
158
|
-
# K5 不传 df 与传 df 的结果不同(否则 A 组可能只是恒真)
|
|
159
|
-
ok(_mdcg.lexical_sim(qb, nb, df=df, n_docs=41) != _mdcg.lexical_sim(qb, nb),
|
|
160
|
-
"K5 传 df 确实改变分数(A 组读数非恒真)",
|
|
161
|
-
(_mdcg.lexical_sim(qb, nb, df=df, n_docs=41), _mdcg.lexical_sim(qb, nb)))
|
|
162
|
-
|
|
163
|
-
|
|
164
|
-
# ---------------------------------------------------------------- E 判据单点
|
|
165
|
-
|
|
166
|
-
def group_e():
|
|
167
|
-
print("== E 判据单点:idf 算术唯一 / df 只在本批统计 / 透传在位 ==")
|
|
168
|
-
src = open(os.path.join(_HERE, "mdcg.py"), encoding="utf-8").read()
|
|
169
|
-
ok(src.count("def idf_of(") == 1, "E1 idf_of 在 mdcg.py 恰定义一次")
|
|
170
|
-
ok(len(re.findall(r"log1p\(\(n_docs - df \+ 0\.5\) / \(df \+ 0\.5\)\)", src)) == 1,
|
|
171
|
-
"E2 IDF 算术在 mdcg.py 恰出现一次(无第二份副本)")
|
|
172
|
-
ok("df=_df, n_docs=_n_docs" in src,
|
|
173
|
-
"E3 _score 把本批 df 透传给 lexical_sim(缺此透传 ⇒ IDF 不生效)")
|
|
174
|
-
ok("for _nb in _nbs:" in src and "_df.get(_g, 0) + 1" in src,
|
|
175
|
-
"E4 df 统计面=本批候选文档(不扫全库、不读盘)")
|
|
176
|
-
bm = open(os.path.join(_HERE, "bm25.py"), encoding="utf-8").read()
|
|
177
|
-
ok("(count - len(self.post[term]) + 0.5) / (len(self.post[term]) + 0.5)" in bm,
|
|
178
|
-
"E5 与 bm25.py 的 IDF 算术同源(同一公式,两处同形)")
|
|
179
|
-
|
|
180
|
-
|
|
181
|
-
# ---------------------------------------------------------------- 变异
|
|
182
|
-
|
|
183
|
-
_MUTATIONS = (
|
|
184
|
-
("M1 撤 _score 的 df 透传(IDF 不再生效)",
|
|
185
|
-
" sim = lexical_sim(qb, nb, mode, df=_df, n_docs=_n_docs)",
|
|
186
|
-
" sim = lexical_sim(qb, nb, mode)",
|
|
187
|
-
("A1", "A2", "A3")),
|
|
188
|
-
("M2 撤 lexical_sim 的 IDF 分支(回退旧公式)",
|
|
189
|
-
" if df is not None and n_docs:",
|
|
190
|
-
" if False: # MUT",
|
|
191
|
-
("A1", "A2", "A3", "K5")),
|
|
192
|
-
("M3 idf 打成常量 1.0(有分支但无判别力 ⇒ 退化为旧公式)",
|
|
193
|
-
" return math.log1p((n_docs - df + 0.5) / (df + 0.5))",
|
|
194
|
-
" return 1.0 # MUT",
|
|
195
|
-
("A1", "A2", "A3")),
|
|
196
|
-
)
|
|
197
|
-
|
|
198
|
-
|
|
199
|
-
def _mutant(patches, name):
|
|
200
|
-
src = open(os.path.join(_HERE, "mdcg.py"), encoding="utf-8").read()
|
|
201
|
-
for a, b in patches:
|
|
202
|
-
if a not in src:
|
|
203
|
-
print("ANCHOR-MISS: %r" % a[:70])
|
|
204
|
-
return None
|
|
205
|
-
src = src.replace(a, b, 1)
|
|
206
|
-
mod = types.ModuleType("md_cg._mdcg_" + name)
|
|
207
|
-
mod.__package__, mod.__file__ = "md_cg", os.path.join(_HERE, "mdcg.py")
|
|
208
|
-
exec(compile(src, "<mutant:%s>" % name, "exec"), mod.__dict__) # noqa: S102
|
|
209
|
-
return mod
|
|
210
|
-
|
|
211
|
-
|
|
212
|
-
def group_mutate():
|
|
213
|
-
print("== 变异:修复回退 ⇒ 指定项必须转红 ==")
|
|
214
|
-
base = _fixture_readout(_mdcg.MdCG, os.path.join(_TMP, "m_base"))
|
|
215
|
-
base_ok = {"A1": base["rank"] == 1,
|
|
216
|
-
"A2": base["rare"] - base["common"] > 1e-3,
|
|
217
|
-
"A3": base["levels"] >= 2,
|
|
218
|
-
"K5": _mdcg.lexical_sim({"甲甲", "乙乙", "甲乙"}, {"甲甲", "甲乙"},
|
|
219
|
-
df={"甲甲": 40, "乙乙": 1, "甲乙": 40},
|
|
220
|
-
n_docs=41)
|
|
221
|
-
!= _mdcg.lexical_sim({"甲甲", "乙乙", "甲乙"}, {"甲甲", "甲乙"})}
|
|
222
|
-
bad = 0
|
|
223
|
-
for title, a, b, expect in _MUTATIONS:
|
|
224
|
-
mod = _mutant([(a, b)], title.split()[0])
|
|
225
|
-
if mod is None:
|
|
226
|
-
bad += 1
|
|
227
|
-
continue
|
|
228
|
-
reds = set()
|
|
229
|
-
try:
|
|
230
|
-
if "A" in "".join(expect):
|
|
231
|
-
r = _fixture_readout(mod.MdCG, os.path.join(
|
|
232
|
-
_TMP, "m_" + title.split()[0]))
|
|
233
|
-
if not (r["rank"] == 1):
|
|
234
|
-
reds.add("A1")
|
|
235
|
-
if not (r["rare"] - r["common"] > 1e-3):
|
|
236
|
-
reds.add("A2")
|
|
237
|
-
if not (r["levels"] >= 2):
|
|
238
|
-
reds.add("A3")
|
|
239
|
-
if "K5" in expect:
|
|
240
|
-
if mod.lexical_sim({"甲甲", "乙乙", "甲乙"}, {"甲甲", "甲乙"},
|
|
241
|
-
df={"甲甲": 40, "乙乙": 1, "甲乙": 40},
|
|
242
|
-
n_docs=41) == mod.lexical_sim(
|
|
243
|
-
{"甲甲", "乙乙", "甲乙"}, {"甲甲", "甲乙"}):
|
|
244
|
-
reds.add("K5")
|
|
245
|
-
except Exception as exc: # 变异导致崩溃也算转红(指名到项)
|
|
246
|
-
print(" 变异 %s 崩溃:%r" % (title, exc))
|
|
247
|
-
reds |= set(expect)
|
|
248
|
-
print(" %s → 红项=%s(预期 %s)" % (title, sorted(reds), list(expect)))
|
|
249
|
-
ok(reds == set(expect) and bool(reds),
|
|
250
|
-
"变异 %s 的红项集合与预期逐项相符" % title.split()[0],
|
|
251
|
-
sorted(reds))
|
|
252
|
-
bad += 0 if reds == set(expect) else 1
|
|
253
|
-
ok(all(base_ok.values()), "M0 正向基线自证(变异前的读数确实成立)", base_ok)
|
|
254
|
-
return bad
|
|
255
|
-
|
|
256
|
-
|
|
257
|
-
def main():
|
|
258
|
-
print("=== issue #363 守卫:缺省 lexical 路的真实打分(IDF)===")
|
|
259
|
-
mutate = "--mutate" in sys.argv
|
|
260
|
-
group_a()
|
|
261
|
-
group_k()
|
|
262
|
-
group_e()
|
|
263
|
-
if mutate:
|
|
264
|
-
bad = group_mutate()
|
|
265
|
-
if bad:
|
|
266
|
-
FAIL.append("变异段 %d 条不合格" % bad)
|
|
267
|
-
print("\n通过 %d / 失败 %d" % (len(PASS), len(FAIL)))
|
|
268
|
-
if FAIL:
|
|
269
|
-
print("失败项:" + ";".join(FAIL))
|
|
270
|
-
return 1
|
|
271
|
-
return 0
|
|
272
|
-
|
|
273
|
-
|
|
274
|
-
if __name__ == "__main__":
|
|
275
|
-
sys.exit(main())
|
|
1
|
+
# -*- coding: utf-8 -*-
|
|
2
|
+
# 功能名:缺省 lexical 路的真实打分(IDF 加权覆盖度,issue #363)
|
|
3
|
+
# 生效条件:python -X utf8 -m md_cg.test_issue363_lexical_idf [--mutate];被测面=md_cg/mdcg.py 的 idf_of / lexical_sim(df,n_docs) / MdCG._score 的 df 统计与透传
|
|
4
|
+
# 子功能:A 组=缺陷行为面(同分退化 → 稀有 bigram 命中者居首、分数可分辨);K 组=非恒真对照(不传 df 的旧公式逐位不变、jaccard 不受影响、idf 与 bm25 同公式且随 df 单调降、同内容文档不产生虚假分差);E 组=判据单点(idf 算术全仓唯一、df 只在 _score 统计、透传在位)
|
|
5
|
+
# 执行:--mutate 时把修复逐条回退(撤 _score 的 df 透传 / 撤 lexical_sim 的 IDF 分支 / 把 idf 打成常量),守卫必须按指定项转红
|
|
6
|
+
# 验证方式:正向全部 PASS 且退出码 0;--mutate 下红项集合须与预期逐项相符(空集合即不合格),变异锚点未命中 ⇒ 退出码 2(fail-closed)
|
|
7
|
+
# 不适用条件:① 不判 Rust 侧(rust/src/retrieval.rs 的 legacy 公式未同步改,跨侧 rank_parity 由 scripts/rank_parity.py 的 jaccard 口径钉住,本件不覆盖);② 不判 `_score` 之外的二文档比对(S5 负记忆抑制仍走不传 df 的旧公式,K1 只钉其不变);③ 不判跨候选池的分数可比性(df 统计面=本批 docs,跨批分数不可横比,已在 lexical_sim docstring 登记);④ 不判冻结集 hit@k/MRR 的绝对量级(随语料浮动,只登记在报告里,不作断言)。
|
|
8
|
+
"""issue #363 守卫:缺省 lexical 路的真实打分(IDF 加权覆盖度)。
|
|
9
|
+
|
|
10
|
+
病象(修前,本守卫 A 组逐条复现):
|
|
11
|
+
查询「甲甲乙乙」在 40 条只含「甲甲」的干扰 + 1 条只含「乙乙」的稀有文档上,
|
|
12
|
+
legacy 口径 `|q∩d|/|q|` 对 41 条文档全部给 0.333333 —— 分数只有 k/|q| 几个
|
|
13
|
+
离散档,名次由 `(-score, -importance, id)` 的 id 兜底决定=写入序:稀有命中
|
|
14
|
+
的 z_rare 排第 41(探针实测 41/41,分数档数 1/41)。
|
|
15
|
+
|
|
16
|
+
修法(判据单点,不新造第二套分词口径):
|
|
17
|
+
· `idf_of(df, n_docs) = log1p((N-df+0.5)/(df+0.5))`——与 `md_cg/bm25.py`
|
|
18
|
+
的 BM25 IDF 同一算术,全仓唯一实现;
|
|
19
|
+
· `lexical_sim(..., df=, n_docs=)` 在 legacy 口径下走
|
|
20
|
+
`Σ_{g∈q∩d} idf(g) / Σ_{g∈q} idf(g)`;**不传 df 时逐位保持旧公式**;
|
|
21
|
+
· `MdCG._score` 按本批候选文档统计查询 bigram 的 df 并透传(与 bm25 的
|
|
22
|
+
posting 面同为「当前候选池」口径,不扫全库)。
|
|
23
|
+
|
|
24
|
+
判别力来源(非恒真):A1 修前 41 → 修后 1;A2 稀有分/常见分差 > 1e-3;
|
|
25
|
+
K1 旧公式逐位相等(防止「到处都变了」);K4 同内容文档同分(防止「凭空分差」);
|
|
26
|
+
E1/E4 单一实现(防止第二份 IDF 公式漂移)。
|
|
27
|
+
"""
|
|
28
|
+
from __future__ import annotations
|
|
29
|
+
|
|
30
|
+
import math
|
|
31
|
+
import os
|
|
32
|
+
import re
|
|
33
|
+
import sys
|
|
34
|
+
import tempfile
|
|
35
|
+
import types
|
|
36
|
+
|
|
37
|
+
sys.stdout.reconfigure(encoding="utf-8", errors="replace")
|
|
38
|
+
|
|
39
|
+
_HERE = os.path.dirname(os.path.abspath(__file__))
|
|
40
|
+
_REPO = os.path.dirname(_HERE)
|
|
41
|
+
if _REPO not in sys.path:
|
|
42
|
+
sys.path.insert(0, _REPO)
|
|
43
|
+
|
|
44
|
+
_TMP = tempfile.mkdtemp(prefix="iss363_")
|
|
45
|
+
os.environ["MDCG_ROOT"] = os.path.join(_TMP, "root")
|
|
46
|
+
os.environ["MDCG_AUX_ROOT"] = os.path.join(_TMP, "auxroot")
|
|
47
|
+
os.environ["MDCG_MASTER_KEY"] = "cd" * 32
|
|
48
|
+
os.environ["MDCG_TOKEN_FILE"] = os.path.join(_TMP, "tokens.json")
|
|
49
|
+
for _k in ("MDCG_TOKEN", "MDCG_VERIFY_KEY", "MDCG_SCORE_MODE", "MDCG_UNIFY_QUERY",
|
|
50
|
+
"MDCG_CN_GRAMS", "MDCG_READ_CACHE"):
|
|
51
|
+
os.environ.pop(_k, None)
|
|
52
|
+
|
|
53
|
+
from md_cg import mdcg as _mdcg # noqa: E402
|
|
54
|
+
|
|
55
|
+
PASS: list = []
|
|
56
|
+
FAIL: list = []
|
|
57
|
+
|
|
58
|
+
|
|
59
|
+
def ok(cond, msg, detail=""):
|
|
60
|
+
(PASS if cond else FAIL).append(msg)
|
|
61
|
+
print((" PASS " if cond else " FAIL ") + msg
|
|
62
|
+
+ ((" ← " + str(detail)) if (detail and not cond) else ""))
|
|
63
|
+
return bool(cond)
|
|
64
|
+
|
|
65
|
+
|
|
66
|
+
# ---------------------------------------------------------------- 夹具
|
|
67
|
+
|
|
68
|
+
QUERY = "甲甲乙乙"
|
|
69
|
+
COMMON = "甲甲 记录 第%d条"
|
|
70
|
+
RARE = "乙乙 独特 内容"
|
|
71
|
+
|
|
72
|
+
|
|
73
|
+
def _build(cg_cls, root, n_common=40):
|
|
74
|
+
cg = cg_cls(root)
|
|
75
|
+
for i in range(n_common):
|
|
76
|
+
cg.add("a_common%02d" % i, COMMON % i, layer="knowledge")
|
|
77
|
+
cg.add("z_rare", RARE, layer="knowledge")
|
|
78
|
+
cg.flush()
|
|
79
|
+
return cg
|
|
80
|
+
|
|
81
|
+
|
|
82
|
+
def _search_rows(cg, k=41):
|
|
83
|
+
res, meta = cg.search(QUERY, k=k, judge=False, record=False)
|
|
84
|
+
return [(n["id"], float(s)) for n, s, _q in res], meta
|
|
85
|
+
|
|
86
|
+
|
|
87
|
+
def _fixture_readout(cg_cls, root):
|
|
88
|
+
cg = _build(cg_cls, root)
|
|
89
|
+
rows, meta = _search_rows(cg)
|
|
90
|
+
ids = [i for i, _ in rows]
|
|
91
|
+
scores = [s for _, s in rows]
|
|
92
|
+
rank = ids.index("z_rare") + 1 if "z_rare" in ids else 0
|
|
93
|
+
rare = scores[ids.index("z_rare")] if rank else 0.0
|
|
94
|
+
common = max((s for i, s in rows if i.startswith("a_common")), default=0.0)
|
|
95
|
+
return {"rank": rank, "rare": rare, "common": common,
|
|
96
|
+
"levels": len({round(s, 12) for s in scores}), "n": len(rows),
|
|
97
|
+
"meta": meta}
|
|
98
|
+
|
|
99
|
+
|
|
100
|
+
# ---------------------------------------------------------------- A 行为面
|
|
101
|
+
|
|
102
|
+
def group_a():
|
|
103
|
+
print("== A 缺陷行为面:同分退化 → 稀有 bigram 命中者居首 ==")
|
|
104
|
+
r = _fixture_readout(_mdcg.MdCG, os.path.join(_TMP, "a1"))
|
|
105
|
+
print(" 读数:rank=%s/%s rare=%.6f common=%.6f levels=%s"
|
|
106
|
+
% (r["rank"], r["n"], r["rare"], r["common"], r["levels"]))
|
|
107
|
+
ok(r["n"] == 41, "A0 夹具非空转:41 条文档全部参与打分(正对照)", r["n"])
|
|
108
|
+
ok(r["rank"] == 1,
|
|
109
|
+
"A1 稀有 bigram(只出现在 1/41 文档)命中者居首(修前 rank=41,写入序兜底)",
|
|
110
|
+
"rank=%s" % r["rank"])
|
|
111
|
+
ok(r["rare"] - r["common"] > 1e-3,
|
|
112
|
+
"A2 稀有/常见命中分数可分辨(修前逐位相同 0.333333)",
|
|
113
|
+
"rare=%r common=%r" % (r["rare"], r["common"]))
|
|
114
|
+
ok(r["levels"] >= 2, "A3 分数档数 > 1(修前 1/41:整批同分)", r["levels"])
|
|
115
|
+
|
|
116
|
+
|
|
117
|
+
# ---------------------------------------------------------------- K 非恒真对照
|
|
118
|
+
|
|
119
|
+
def group_k():
|
|
120
|
+
print("== K 非恒真对照:旧公式 / jaccard / idf 公式 / 无虚假分差 ==")
|
|
121
|
+
qb = {"甲甲", "乙乙", "甲乙"}
|
|
122
|
+
nb = {"甲甲", "甲乙"}
|
|
123
|
+
# K1 不传 df ⇒ 旧公式逐位不变(S5 等二文档比对路径)
|
|
124
|
+
ok(_mdcg.lexical_sim(qb, nb) == len(qb & nb) / len(qb),
|
|
125
|
+
"K1 不传 df 时 legacy 仍是 |q∩d|/|q|(旧公式逐位不变)",
|
|
126
|
+
_mdcg.lexical_sim(qb, nb))
|
|
127
|
+
ok(_mdcg.lexical_sim(set(), nb) == 0.0 and _mdcg.lexical_sim(qb, set()) == 0.0,
|
|
128
|
+
"K1b 空集合恒 0.0(边界不因 IDF 改变)")
|
|
129
|
+
# K2 jaccard 不受 df 影响
|
|
130
|
+
df = {"甲甲": 40, "乙乙": 1, "甲乙": 40}
|
|
131
|
+
ok(_mdcg.lexical_sim(qb, nb, mode="jaccard", df=df, n_docs=41)
|
|
132
|
+
== len(qb & nb) / len(qb | nb),
|
|
133
|
+
"K2 jaccard 口径不因传入 df 改变(口径隔离)")
|
|
134
|
+
# K3 idf 与 bm25 同公式
|
|
135
|
+
for d in (0, 1, 20, 40, 41):
|
|
136
|
+
want = math.log1p((41 - d + 0.5) / (d + 0.5))
|
|
137
|
+
got = _mdcg.idf_of(d, 41)
|
|
138
|
+
ok(abs(got - want) < 1e-15,
|
|
139
|
+
"K3 idf_of(df=%d, N=41) 与 bm25 公式逐位一致" % d, (got, want))
|
|
140
|
+
ok(_mdcg.idf_of(0, 41) > _mdcg.idf_of(20, 41) > _mdcg.idf_of(40, 41),
|
|
141
|
+
"K3b idf 随 df 单调递减(高频 bigram 权重更低)")
|
|
142
|
+
ok(_mdcg.idf_of(5, 0) == 0.0, "K3c N=0 时 idf 恒 0(不猜测)")
|
|
143
|
+
# K4 同内容文档不得产生虚假分差
|
|
144
|
+
ok(_mdcg.lexical_sim(qb, nb, df=df, n_docs=41)
|
|
145
|
+
== _mdcg.lexical_sim(qb, nb, df=df, n_docs=41),
|
|
146
|
+
"K4 同一输入同分(打分是纯函数)")
|
|
147
|
+
root = os.path.join(_TMP, "k4")
|
|
148
|
+
cg = _mdcg.MdCG(root)
|
|
149
|
+
for i in range(6):
|
|
150
|
+
cg.add("dup%d" % i, "丙丙 同样内容", layer="knowledge")
|
|
151
|
+
cg.flush()
|
|
152
|
+
# 用与内容同词的查询(有 LIKE 命中 ⇒ 走真实打分路径,而不是 0 分兜底池)
|
|
153
|
+
rows = [(n["id"], float(s)) for n, s, _q in
|
|
154
|
+
cg.search("丙丙", k=6, judge=False, record=False)[0]]
|
|
155
|
+
ok(len(rows) == 6 and len({round(s, 12) for _, s in rows}) == 1,
|
|
156
|
+
"K4b 同内容文档分数逐位相同(IDF 不制造虚假分差)",
|
|
157
|
+
[s for _, s in rows])
|
|
158
|
+
# K5 不传 df 与传 df 的结果不同(否则 A 组可能只是恒真)
|
|
159
|
+
ok(_mdcg.lexical_sim(qb, nb, df=df, n_docs=41) != _mdcg.lexical_sim(qb, nb),
|
|
160
|
+
"K5 传 df 确实改变分数(A 组读数非恒真)",
|
|
161
|
+
(_mdcg.lexical_sim(qb, nb, df=df, n_docs=41), _mdcg.lexical_sim(qb, nb)))
|
|
162
|
+
|
|
163
|
+
|
|
164
|
+
# ---------------------------------------------------------------- E 判据单点
|
|
165
|
+
|
|
166
|
+
def group_e():
|
|
167
|
+
print("== E 判据单点:idf 算术唯一 / df 只在本批统计 / 透传在位 ==")
|
|
168
|
+
src = open(os.path.join(_HERE, "mdcg.py"), encoding="utf-8").read()
|
|
169
|
+
ok(src.count("def idf_of(") == 1, "E1 idf_of 在 mdcg.py 恰定义一次")
|
|
170
|
+
ok(len(re.findall(r"log1p\(\(n_docs - df \+ 0\.5\) / \(df \+ 0\.5\)\)", src)) == 1,
|
|
171
|
+
"E2 IDF 算术在 mdcg.py 恰出现一次(无第二份副本)")
|
|
172
|
+
ok("df=_df, n_docs=_n_docs" in src,
|
|
173
|
+
"E3 _score 把本批 df 透传给 lexical_sim(缺此透传 ⇒ IDF 不生效)")
|
|
174
|
+
ok("for _nb in _nbs:" in src and "_df.get(_g, 0) + 1" in src,
|
|
175
|
+
"E4 df 统计面=本批候选文档(不扫全库、不读盘)")
|
|
176
|
+
bm = open(os.path.join(_HERE, "bm25.py"), encoding="utf-8").read()
|
|
177
|
+
ok("(count - len(self.post[term]) + 0.5) / (len(self.post[term]) + 0.5)" in bm,
|
|
178
|
+
"E5 与 bm25.py 的 IDF 算术同源(同一公式,两处同形)")
|
|
179
|
+
|
|
180
|
+
|
|
181
|
+
# ---------------------------------------------------------------- 变异
|
|
182
|
+
|
|
183
|
+
_MUTATIONS = (
|
|
184
|
+
("M1 撤 _score 的 df 透传(IDF 不再生效)",
|
|
185
|
+
" sim = lexical_sim(qb, nb, mode, df=_df, n_docs=_n_docs)",
|
|
186
|
+
" sim = lexical_sim(qb, nb, mode)",
|
|
187
|
+
("A1", "A2", "A3")),
|
|
188
|
+
("M2 撤 lexical_sim 的 IDF 分支(回退旧公式)",
|
|
189
|
+
" if df is not None and n_docs:",
|
|
190
|
+
" if False: # MUT",
|
|
191
|
+
("A1", "A2", "A3", "K5")),
|
|
192
|
+
("M3 idf 打成常量 1.0(有分支但无判别力 ⇒ 退化为旧公式)",
|
|
193
|
+
" return math.log1p((n_docs - df + 0.5) / (df + 0.5))",
|
|
194
|
+
" return 1.0 # MUT",
|
|
195
|
+
("A1", "A2", "A3")),
|
|
196
|
+
)
|
|
197
|
+
|
|
198
|
+
|
|
199
|
+
def _mutant(patches, name):
|
|
200
|
+
src = open(os.path.join(_HERE, "mdcg.py"), encoding="utf-8").read()
|
|
201
|
+
for a, b in patches:
|
|
202
|
+
if a not in src:
|
|
203
|
+
print("ANCHOR-MISS: %r" % a[:70])
|
|
204
|
+
return None
|
|
205
|
+
src = src.replace(a, b, 1)
|
|
206
|
+
mod = types.ModuleType("md_cg._mdcg_" + name)
|
|
207
|
+
mod.__package__, mod.__file__ = "md_cg", os.path.join(_HERE, "mdcg.py")
|
|
208
|
+
exec(compile(src, "<mutant:%s>" % name, "exec"), mod.__dict__) # noqa: S102
|
|
209
|
+
return mod
|
|
210
|
+
|
|
211
|
+
|
|
212
|
+
def group_mutate():
|
|
213
|
+
print("== 变异:修复回退 ⇒ 指定项必须转红 ==")
|
|
214
|
+
base = _fixture_readout(_mdcg.MdCG, os.path.join(_TMP, "m_base"))
|
|
215
|
+
base_ok = {"A1": base["rank"] == 1,
|
|
216
|
+
"A2": base["rare"] - base["common"] > 1e-3,
|
|
217
|
+
"A3": base["levels"] >= 2,
|
|
218
|
+
"K5": _mdcg.lexical_sim({"甲甲", "乙乙", "甲乙"}, {"甲甲", "甲乙"},
|
|
219
|
+
df={"甲甲": 40, "乙乙": 1, "甲乙": 40},
|
|
220
|
+
n_docs=41)
|
|
221
|
+
!= _mdcg.lexical_sim({"甲甲", "乙乙", "甲乙"}, {"甲甲", "甲乙"})}
|
|
222
|
+
bad = 0
|
|
223
|
+
for title, a, b, expect in _MUTATIONS:
|
|
224
|
+
mod = _mutant([(a, b)], title.split()[0])
|
|
225
|
+
if mod is None:
|
|
226
|
+
bad += 1
|
|
227
|
+
continue
|
|
228
|
+
reds = set()
|
|
229
|
+
try:
|
|
230
|
+
if "A" in "".join(expect):
|
|
231
|
+
r = _fixture_readout(mod.MdCG, os.path.join(
|
|
232
|
+
_TMP, "m_" + title.split()[0]))
|
|
233
|
+
if not (r["rank"] == 1):
|
|
234
|
+
reds.add("A1")
|
|
235
|
+
if not (r["rare"] - r["common"] > 1e-3):
|
|
236
|
+
reds.add("A2")
|
|
237
|
+
if not (r["levels"] >= 2):
|
|
238
|
+
reds.add("A3")
|
|
239
|
+
if "K5" in expect:
|
|
240
|
+
if mod.lexical_sim({"甲甲", "乙乙", "甲乙"}, {"甲甲", "甲乙"},
|
|
241
|
+
df={"甲甲": 40, "乙乙": 1, "甲乙": 40},
|
|
242
|
+
n_docs=41) == mod.lexical_sim(
|
|
243
|
+
{"甲甲", "乙乙", "甲乙"}, {"甲甲", "甲乙"}):
|
|
244
|
+
reds.add("K5")
|
|
245
|
+
except Exception as exc: # 变异导致崩溃也算转红(指名到项)
|
|
246
|
+
print(" 变异 %s 崩溃:%r" % (title, exc))
|
|
247
|
+
reds |= set(expect)
|
|
248
|
+
print(" %s → 红项=%s(预期 %s)" % (title, sorted(reds), list(expect)))
|
|
249
|
+
ok(reds == set(expect) and bool(reds),
|
|
250
|
+
"变异 %s 的红项集合与预期逐项相符" % title.split()[0],
|
|
251
|
+
sorted(reds))
|
|
252
|
+
bad += 0 if reds == set(expect) else 1
|
|
253
|
+
ok(all(base_ok.values()), "M0 正向基线自证(变异前的读数确实成立)", base_ok)
|
|
254
|
+
return bad
|
|
255
|
+
|
|
256
|
+
|
|
257
|
+
def main():
|
|
258
|
+
print("=== issue #363 守卫:缺省 lexical 路的真实打分(IDF)===")
|
|
259
|
+
mutate = "--mutate" in sys.argv
|
|
260
|
+
group_a()
|
|
261
|
+
group_k()
|
|
262
|
+
group_e()
|
|
263
|
+
if mutate:
|
|
264
|
+
bad = group_mutate()
|
|
265
|
+
if bad:
|
|
266
|
+
FAIL.append("变异段 %d 条不合格" % bad)
|
|
267
|
+
print("\n通过 %d / 失败 %d" % (len(PASS), len(FAIL)))
|
|
268
|
+
if FAIL:
|
|
269
|
+
print("失败项:" + ";".join(FAIL))
|
|
270
|
+
return 1
|
|
271
|
+
return 0
|
|
272
|
+
|
|
273
|
+
|
|
274
|
+
if __name__ == "__main__":
|
|
275
|
+
sys.exit(main())
|