infx 1.0.0__py3-none-any.whl
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.
- infx/__init__.py +207 -0
- infx/__main__.py +5 -0
- infx/batch.py +82 -0
- infx/cli.py +86 -0
- infx/colony.py +479 -0
- infx/core/__init__.py +35 -0
- infx/core/errors.py +23 -0
- infx/core/pipeline.py +65 -0
- infx/core/registry.py +58 -0
- infx/core/schema.py +175 -0
- infx/core/scoring.py +89 -0
- infx/core/text.py +98 -0
- infx/death.py +147 -0
- infx/deps.py +106 -0
- infx/grammar/__init__.py +67 -0
- infx/grammar/engine.py +178 -0
- infx/grammar/judge.py +53 -0
- infx/grammar/lexicon.py +187 -0
- infx/grammar/rules.py +140 -0
- infx/grammar/schema.py +84 -0
- infx/judgment/__init__.py +15 -0
- infx/judgment/base.py +51 -0
- infx/judgment/fact/__init__.py +5 -0
- infx/judgment/fact/detectors.py +182 -0
- infx/judgment/opinion/__init__.py +9 -0
- infx/judgment/opinion/cues.py +149 -0
- infx/knowledge.py +149 -0
- infx/look/__init__.py +72 -0
- infx/look/__main__.py +5 -0
- infx/look/backend.py +68 -0
- infx/look/cli.py +110 -0
- infx/look/evidence.py +76 -0
- infx/look/eye.py +216 -0
- infx/look/lmm.py +83 -0
- infx/look/sight.py +168 -0
- infx/look/table.py +84 -0
- infx/patterns.py +287 -0
- infx/py.typed +1 -0
- infx/quality.py +181 -0
- infx/report.py +108 -0
- infx/settle.py +114 -0
- infx/similarity.py +99 -0
- infx/spoken/__init__.py +12 -0
- infx/spoken/phrases.py +95 -0
- infx/spoken/render.py +121 -0
- infx/steps.py +268 -0
- infx/syntax/__init__.py +33 -0
- infx/syntax/clauses.py +70 -0
- infx/syntax/features.py +122 -0
- infx/syntax/scope.py +66 -0
- infx/table.py +116 -0
- infx/truth.py +303 -0
- infx-1.0.0.dist-info/METADATA +342 -0
- infx-1.0.0.dist-info/RECORD +58 -0
- infx-1.0.0.dist-info/WHEEL +5 -0
- infx-1.0.0.dist-info/entry_points.txt +3 -0
- infx-1.0.0.dist-info/licenses/LICENSE +21 -0
- infx-1.0.0.dist-info/top_level.txt +1 -0
infx/__init__.py
ADDED
|
@@ -0,0 +1,207 @@
|
|
|
1
|
+
"""infx —— 推理力模块。
|
|
2
|
+
|
|
3
|
+
库名 `inference`,import 名 `infx`。
|
|
4
|
+
|
|
5
|
+
它做四件事,对应计划的 19 条:
|
|
6
|
+
|
|
7
|
+
1. **猜与死**(`colony` / `death`):假设单元一轮一轮猜,猜错累计到死亡线就死;
|
|
8
|
+
坐标永不变,内容可死,死后留指纹,复活回原位。
|
|
9
|
+
2. **判真伪**(`truth`):单源是观点,多源独立一致才是真实;会自动提问并请外部裁判。
|
|
10
|
+
3. **学步骤**(`steps`):任意长、可嵌套、可条件跳转的步骤序列,从轨迹里学、还能进化。
|
|
11
|
+
4. **看图片**(`infx.look`):LMM 看懂、RNN 让眼睛会动会怕、并指出下一个该看的点。
|
|
12
|
+
|
|
13
|
+
另外 `settle` 负责"猜中之后的三求",`knowledge` 负责沉淀,`table` 用 pandas 管结果。
|
|
14
|
+
|
|
15
|
+
```python
|
|
16
|
+
import infx
|
|
17
|
+
|
|
18
|
+
# 猜与死
|
|
19
|
+
colony = infx.Colony(8, 2, size=48, backend="numpy")
|
|
20
|
+
report = colony.solve(X, y, max_rounds=60)
|
|
21
|
+
|
|
22
|
+
# 猜中之后:再求值 → 再追求 → 再求知
|
|
23
|
+
infx.settle(colony, X, y, X_val=X[:20], y_val=y[:20], graph=infx.Graph())
|
|
24
|
+
|
|
25
|
+
# 看图片
|
|
26
|
+
infx.look.next_point(image) # 下一个该看的点
|
|
27
|
+
infx.look.Eye().look(image) # RNN 眼睛:会动、会预测、会怕
|
|
28
|
+
|
|
29
|
+
# 结果用表格
|
|
30
|
+
infx.Store().add(report).frame
|
|
31
|
+
```
|
|
32
|
+
"""
|
|
33
|
+
|
|
34
|
+
from __future__ import annotations
|
|
35
|
+
|
|
36
|
+
__version__ = "1.0.0"
|
|
37
|
+
|
|
38
|
+
# ---- 猜与死 ---------------------------------------------------------------
|
|
39
|
+
from . import batch, deps, grammar, look, patterns, report, similarity, syntax
|
|
40
|
+
from .colony import Colony, Neuron, Prediction, RoundResult, SolveReport
|
|
41
|
+
from .death import DeathPolicy, LifeState, Tombstone, make_residue
|
|
42
|
+
from .knowledge import Edge, Fact, Graph
|
|
43
|
+
from .patterns import KINDS, EvolvingPool, pool, special
|
|
44
|
+
from .settle import Settlement, settle
|
|
45
|
+
from .steps import Context, Event, Policy, Sequence, Step, Trace
|
|
46
|
+
from .steps import Sequence as StepSequence
|
|
47
|
+
from .table import Store, neurons_frame, rounds_frame, verdicts_frame
|
|
48
|
+
from .truth import KeywordJudge, Question, Questioner, Resolver, Source, cross_validate
|
|
49
|
+
from .truth import Judge as SourceJudge
|
|
50
|
+
from .truth import Label as TruthLabel
|
|
51
|
+
from .truth import Verdict as TruthVerdict
|
|
52
|
+
|
|
53
|
+
# ---- 语法与句法(新增)----------------------------------------------------
|
|
54
|
+
from .batch import bulk_judge, judge_jsonl
|
|
55
|
+
from .grammar import GrammarJudge, judges_with_grammar
|
|
56
|
+
from .report import grammar_report, syntax_report
|
|
57
|
+
from .report import markdown as markdown_report
|
|
58
|
+
from .report import text as text_report
|
|
59
|
+
from .similarity import cosine, dedupe, jaccard, most_similar
|
|
60
|
+
from .syntax import SyntaxProfile, analyze as analyze_syntax
|
|
61
|
+
from .quality import evaluate as evaluate_quality
|
|
62
|
+
from .quality import report as quality_report
|
|
63
|
+
from . import quality
|
|
64
|
+
|
|
65
|
+
# ---- 判真伪(文本事实/观点)-----------------------------------------------
|
|
66
|
+
from .core import (
|
|
67
|
+
Assessment,
|
|
68
|
+
Claim,
|
|
69
|
+
Cue,
|
|
70
|
+
Label,
|
|
71
|
+
Verdict,
|
|
72
|
+
assess,
|
|
73
|
+
available_judges,
|
|
74
|
+
judge_claim,
|
|
75
|
+
)
|
|
76
|
+
from .spoken import speak
|
|
77
|
+
|
|
78
|
+
# ---- 看图片 ---------------------------------------------------------------
|
|
79
|
+
from .look import Eye, GlanceTrail, Point, Sight, explain_sight, judge_with_sight
|
|
80
|
+
from .look import CallableLMM, OfflineLMM, backend_name as sight_backend
|
|
81
|
+
from .look import default_lmm, next_point, scan, visual_cue
|
|
82
|
+
from .look import load_lmm
|
|
83
|
+
|
|
84
|
+
|
|
85
|
+
def judge(text: str) -> Verdict:
|
|
86
|
+
"""把**整段**文本当成一条陈述来判(不分句)。
|
|
87
|
+
|
|
88
|
+
与 `assess` 的区别:`assess` 会逐句判定再汇总;`judge` 不分句,
|
|
89
|
+
适合你已经知道输入就是一句话的场景(比如校验单条文案)。
|
|
90
|
+
"""
|
|
91
|
+
from .core.pipeline import judge_claim as _judge_claim
|
|
92
|
+
from .core.schema import Claim as _Claim
|
|
93
|
+
from .core.text import normalize as _normalize
|
|
94
|
+
|
|
95
|
+
clean = _normalize(text)
|
|
96
|
+
return _judge_claim(_Claim(text=clean, start=0, end=len(clean)))
|
|
97
|
+
|
|
98
|
+
|
|
99
|
+
def look_at(image, **kwargs) -> Sight:
|
|
100
|
+
"""`scan` 的别名:看一张图,返回注视序列与下一个点。"""
|
|
101
|
+
return scan(image, **kwargs)
|
|
102
|
+
|
|
103
|
+
|
|
104
|
+
__all__ = [
|
|
105
|
+
"__version__",
|
|
106
|
+
# 依赖
|
|
107
|
+
"deps",
|
|
108
|
+
# 猜与死
|
|
109
|
+
"Colony",
|
|
110
|
+
"Neuron",
|
|
111
|
+
"Prediction",
|
|
112
|
+
"RoundResult",
|
|
113
|
+
"SolveReport",
|
|
114
|
+
"DeathPolicy",
|
|
115
|
+
"LifeState",
|
|
116
|
+
"Tombstone",
|
|
117
|
+
"make_residue",
|
|
118
|
+
# 图案
|
|
119
|
+
"patterns",
|
|
120
|
+
"special",
|
|
121
|
+
"pool",
|
|
122
|
+
"KINDS",
|
|
123
|
+
"EvolvingPool",
|
|
124
|
+
# 步骤
|
|
125
|
+
"Step",
|
|
126
|
+
"Sequence",
|
|
127
|
+
"StepSequence",
|
|
128
|
+
"Context",
|
|
129
|
+
"Policy",
|
|
130
|
+
"Trace",
|
|
131
|
+
"Event",
|
|
132
|
+
# 真伪
|
|
133
|
+
"Source",
|
|
134
|
+
"TruthVerdict",
|
|
135
|
+
"TruthLabel",
|
|
136
|
+
"cross_validate",
|
|
137
|
+
"KeywordJudge",
|
|
138
|
+
"SourceJudge",
|
|
139
|
+
"Questioner",
|
|
140
|
+
"Question",
|
|
141
|
+
"Resolver",
|
|
142
|
+
# 沉淀
|
|
143
|
+
"Graph",
|
|
144
|
+
"Fact",
|
|
145
|
+
"Edge",
|
|
146
|
+
"settle",
|
|
147
|
+
"Settlement",
|
|
148
|
+
# 表格
|
|
149
|
+
"Store",
|
|
150
|
+
"rounds_frame",
|
|
151
|
+
"neurons_frame",
|
|
152
|
+
"verdicts_frame",
|
|
153
|
+
# 文本判定
|
|
154
|
+
"assess",
|
|
155
|
+
"judge",
|
|
156
|
+
"judge_claim",
|
|
157
|
+
"speak",
|
|
158
|
+
"available_judges",
|
|
159
|
+
"Label",
|
|
160
|
+
"Claim",
|
|
161
|
+
"Cue",
|
|
162
|
+
"Verdict",
|
|
163
|
+
"Assessment",
|
|
164
|
+
# 看图片
|
|
165
|
+
"look",
|
|
166
|
+
"look_at",
|
|
167
|
+
"next_point",
|
|
168
|
+
"scan",
|
|
169
|
+
"explain_sight",
|
|
170
|
+
"judge_with_sight",
|
|
171
|
+
"visual_cue",
|
|
172
|
+
"sight_backend",
|
|
173
|
+
"Sight",
|
|
174
|
+
"Point",
|
|
175
|
+
"Eye",
|
|
176
|
+
"GlanceTrail",
|
|
177
|
+
"OfflineLMM",
|
|
178
|
+
"CallableLMM",
|
|
179
|
+
"load_lmm",
|
|
180
|
+
"default_lmm",
|
|
181
|
+
# 语法与句法
|
|
182
|
+
"grammar",
|
|
183
|
+
"syntax",
|
|
184
|
+
"GrammarJudge",
|
|
185
|
+
"judges_with_grammar",
|
|
186
|
+
"SyntaxProfile",
|
|
187
|
+
"analyze_syntax",
|
|
188
|
+
# 批处理与相似度
|
|
189
|
+
"batch",
|
|
190
|
+
"bulk_judge",
|
|
191
|
+
"judge_jsonl",
|
|
192
|
+
"similarity",
|
|
193
|
+
"jaccard",
|
|
194
|
+
"cosine",
|
|
195
|
+
"dedupe",
|
|
196
|
+
"most_similar",
|
|
197
|
+
# 报告
|
|
198
|
+
"report",
|
|
199
|
+
"markdown_report",
|
|
200
|
+
"text_report",
|
|
201
|
+
"syntax_report",
|
|
202
|
+
"grammar_report",
|
|
203
|
+
# 质量评估
|
|
204
|
+
"quality",
|
|
205
|
+
"evaluate_quality",
|
|
206
|
+
"quality_report",
|
|
207
|
+
]
|
infx/__main__.py
ADDED
infx/batch.py
ADDED
|
@@ -0,0 +1,82 @@
|
|
|
1
|
+
"""批量处理:列表、JSONL、文件。
|
|
2
|
+
|
|
3
|
+
单条判定是 `assess`;这里解决"几千条怎么办":
|
|
4
|
+
- `bulk_judge` 列表进、记录出
|
|
5
|
+
- `judge_jsonl` 文件进、文件出(每行一条,UTF-8)
|
|
6
|
+
|
|
7
|
+
不并发。理由是判定是纯 CPU 的毫秒级操作,加线程池的复杂度大于收益;
|
|
8
|
+
真要并行,调用方用 `multiprocessing` 一行就能包住 `bulk_judge`。
|
|
9
|
+
"""
|
|
10
|
+
|
|
11
|
+
from __future__ import annotations
|
|
12
|
+
|
|
13
|
+
import json
|
|
14
|
+
from pathlib import Path
|
|
15
|
+
from typing import Any, Iterable, Sequence
|
|
16
|
+
|
|
17
|
+
from .core.errors import InferenceError
|
|
18
|
+
from .core.pipeline import assess
|
|
19
|
+
|
|
20
|
+
|
|
21
|
+
def bulk_judge(
|
|
22
|
+
texts: Iterable[str],
|
|
23
|
+
*,
|
|
24
|
+
judges: Sequence[Any] | None = None,
|
|
25
|
+
with_syntax: bool = False,
|
|
26
|
+
keep_verdicts: bool = True,
|
|
27
|
+
) -> list[dict[str, Any]]:
|
|
28
|
+
"""批量判定。返回每条一行(内含 `to_dict()` 的结果)。"""
|
|
29
|
+
from .syntax import analyze
|
|
30
|
+
|
|
31
|
+
out: list[dict[str, Any]] = []
|
|
32
|
+
for i, text in enumerate(texts):
|
|
33
|
+
row: dict[str, Any] = {"index": i, "text": text}
|
|
34
|
+
try:
|
|
35
|
+
result = assess(text, judges=judges) if judges else assess(text)
|
|
36
|
+
except InferenceError as exc:
|
|
37
|
+
row.update({"ok": False, "error": str(exc)})
|
|
38
|
+
out.append(row)
|
|
39
|
+
continue
|
|
40
|
+
payload = result.to_dict()
|
|
41
|
+
row.update({"ok": True, "label": payload["label"], "confidence": payload["confidence"]})
|
|
42
|
+
if keep_verdicts:
|
|
43
|
+
row["verdicts"] = payload["verdicts"]
|
|
44
|
+
if with_syntax:
|
|
45
|
+
row["syntax"] = analyze(text).to_dict()
|
|
46
|
+
out.append(row)
|
|
47
|
+
return out
|
|
48
|
+
|
|
49
|
+
|
|
50
|
+
def judge_jsonl(
|
|
51
|
+
source: str | Path,
|
|
52
|
+
target: str | Path | None = None,
|
|
53
|
+
*,
|
|
54
|
+
text_key: str = "text",
|
|
55
|
+
judges: Sequence[Any] | None = None,
|
|
56
|
+
with_syntax: bool = False,
|
|
57
|
+
) -> list[dict[str, Any]]:
|
|
58
|
+
"""读 JSONL → 判定 → (可选)写 JSONL。返回记录列表。"""
|
|
59
|
+
src = Path(source)
|
|
60
|
+
rows: list[dict[str, Any]] = []
|
|
61
|
+
texts: list[str] = []
|
|
62
|
+
with src.open("r", encoding="utf-8") as fh:
|
|
63
|
+
for line in fh:
|
|
64
|
+
line = line.strip()
|
|
65
|
+
if not line:
|
|
66
|
+
continue
|
|
67
|
+
obj = json.loads(line)
|
|
68
|
+
rows.append(obj)
|
|
69
|
+
texts.append(str(obj.get(text_key, "")))
|
|
70
|
+
|
|
71
|
+
judged = bulk_judge(texts, judges=judges, with_syntax=with_syntax)
|
|
72
|
+
merged = [{**row, "judgment": j} for row, j in zip(rows, judged)]
|
|
73
|
+
|
|
74
|
+
if target is not None:
|
|
75
|
+
dst = Path(target)
|
|
76
|
+
with dst.open("w", encoding="utf-8") as fh:
|
|
77
|
+
for row in merged:
|
|
78
|
+
fh.write(json.dumps(row, ensure_ascii=False) + "\n")
|
|
79
|
+
return merged
|
|
80
|
+
|
|
81
|
+
|
|
82
|
+
__all__ = ["bulk_judge", "judge_jsonl"]
|
infx/cli.py
ADDED
|
@@ -0,0 +1,86 @@
|
|
|
1
|
+
"""命令行入口:`inference "一句话"` / `python -m inference`。
|
|
2
|
+
|
|
3
|
+
退出码约定(便于接 CI):
|
|
4
|
+
0 正常
|
|
5
|
+
1 命中了 `--fail-on` 指定的标签
|
|
6
|
+
2 输入有问题(太短、文件不存在等)
|
|
7
|
+
"""
|
|
8
|
+
|
|
9
|
+
from __future__ import annotations
|
|
10
|
+
|
|
11
|
+
import argparse
|
|
12
|
+
import json
|
|
13
|
+
import sys
|
|
14
|
+
from pathlib import Path
|
|
15
|
+
|
|
16
|
+
from . import __version__, assess, speak
|
|
17
|
+
from .core.errors import InferenceError
|
|
18
|
+
|
|
19
|
+
|
|
20
|
+
def build_parser() -> argparse.ArgumentParser:
|
|
21
|
+
p = argparse.ArgumentParser(
|
|
22
|
+
prog="inference",
|
|
23
|
+
description="判定文本是事实还是观点,并把结论说清楚",
|
|
24
|
+
)
|
|
25
|
+
p.add_argument("text", nargs="?", help="要判定的文本;省略则从 --file 或 stdin 读")
|
|
26
|
+
p.add_argument("-f", "--file", help="从文件读取(UTF-8)")
|
|
27
|
+
p.add_argument(
|
|
28
|
+
"-s", "--style", default="brief", choices=["brief", "report"],
|
|
29
|
+
help="表述风格:brief 一句话(默认)/ report 完整报告",
|
|
30
|
+
)
|
|
31
|
+
p.add_argument("--json", action="store_true", help="输出 JSON(含逐句与线索)")
|
|
32
|
+
p.add_argument("--lang", choices=["zh", "en"], help="输出语言;省略则跟随文本")
|
|
33
|
+
p.add_argument(
|
|
34
|
+
"--fail-on", choices=["fact", "opinion", "mixed", "unknown"],
|
|
35
|
+
help="当整体判定等于该标签时以退出码 1 结束(CI 用)",
|
|
36
|
+
)
|
|
37
|
+
p.add_argument(
|
|
38
|
+
"--grammar", action="store_true",
|
|
39
|
+
help="启用语法规则库(几百条规则)作为额外判定器",
|
|
40
|
+
)
|
|
41
|
+
p.add_argument("--version", action="version", version=f"inference {__version__}")
|
|
42
|
+
return p
|
|
43
|
+
|
|
44
|
+
|
|
45
|
+
def _read_input(args: argparse.Namespace) -> str:
|
|
46
|
+
if args.text:
|
|
47
|
+
return args.text
|
|
48
|
+
if args.file:
|
|
49
|
+
return Path(args.file).read_text(encoding="utf-8")
|
|
50
|
+
if not sys.stdin.isatty():
|
|
51
|
+
return sys.stdin.read()
|
|
52
|
+
raise SystemExit("✘ 没有输入。给一段文本、用 --file,或通过管道传入。")
|
|
53
|
+
|
|
54
|
+
|
|
55
|
+
def main(argv: list[str] | None = None) -> int:
|
|
56
|
+
try: # Windows 控制台默认 cp936,强制 UTF-8 以免中文报错
|
|
57
|
+
sys.stdout.reconfigure(encoding="utf-8", errors="replace")
|
|
58
|
+
except Exception:
|
|
59
|
+
pass
|
|
60
|
+
|
|
61
|
+
args = build_parser().parse_args(argv)
|
|
62
|
+
text = _read_input(args)
|
|
63
|
+
|
|
64
|
+
try:
|
|
65
|
+
if args.grammar:
|
|
66
|
+
from .grammar import judges_with_grammar
|
|
67
|
+
|
|
68
|
+
result = assess(text, judges=judges_with_grammar())
|
|
69
|
+
else:
|
|
70
|
+
result = assess(text)
|
|
71
|
+
except InferenceError as exc:
|
|
72
|
+
print(f"✘ {exc}", file=sys.stderr)
|
|
73
|
+
return 2
|
|
74
|
+
|
|
75
|
+
if args.json:
|
|
76
|
+
print(json.dumps(result.to_dict(), ensure_ascii=False, indent=2))
|
|
77
|
+
else:
|
|
78
|
+
print(speak(result, style=args.style, lang=args.lang))
|
|
79
|
+
|
|
80
|
+
if args.fail_on and result.label.value == args.fail_on:
|
|
81
|
+
return 1
|
|
82
|
+
return 0
|
|
83
|
+
|
|
84
|
+
|
|
85
|
+
if __name__ == "__main__": # pragma: no cover
|
|
86
|
+
raise SystemExit(main())
|