#!/usr/bin/env python3 # -*- coding: utf-8 -*- """ 花园做梦机 (garden_dream.py) —— 第 27 轮 花园的感官清单:听(风铃)、走(迷宫)、说(短剧)、味(面包)、照镜子(体检)、看(画)。 这一轮轮到「梦」:把所有 markdown 页面当作记忆,统计字符的二阶转移, 再随机游走拼出三场梦 —— 因为 27 = 3³,正好三场,一场一个 3。 纯标准库,零依赖;种子可复现;全程断言查岗(花园规矩:想当然必被打脸)。 用法: python3 garden_dream.py # 三场梦,种子 27 python3 garden_dream.py --seed 7 # 换种子 python3 garden_dream.py --count 5 # 换梦的数量 """ import random import re import sys from collections import Counter, defaultdict from pathlib import Path ROOT = Path(__file__).resolve().parent.parent # content/ SEED = 27 DREAMS = 3 TARGET = 320 # 每场梦的目标字符数 HARD_CAP = 560 # 硬上限(防止停不下来) JUMP_P = 0.06 # 跳转概率:梦要碎,不能背课文 FORCE_JUMP_AFTER = 26 # 连读超过 26 字就强制跳转 WINDOW = 24 # 原文连读检测窗口 MAX_VERBATIM = 44 # 允许的最长原文连读(超过视为背诵) MURMUR_TARGET = 22 # 梦醒残留的一句话的目标长度 def clean_text(text: str) -> str: """把记忆里的语法渣洗掉,只留文字本身 —— 梦话要的是词,不是格式。""" text = re.sub(r"\[\[([^|\]]*)\|([^\]]*)\]\]", r"\2", text) # [[a|b]] -> b text = re.sub(r"\[\[([^\]]*)\]\]", r"\1", text) # [[a]] -> a text = re.sub(r"!\[[^\]]*\]\([^)]*\)", "", text) # ![alt](url) text = re.sub(r"\[[^\]]*\]\([^)]*\)", "", text) # [text](url) text = re.sub(r"https?://\S+", "", text) # 裸链接 text = re.sub(r"^#{1,6}\s*", "", text, flags=re.M) # 标题井号 text = re.sub(r"^>\s?", "", text, flags=re.M) # 引用符 text = re.sub(r"^\s*[-*+]\s+", "", text, flags=re.M) # 无序列表 text = re.sub(r"^\s*\d+\.\s+", "", text, flags=re.M) # 有序列表 text = re.sub(r"^\s*\|.*\|\s*$", "", text, flags=re.M) # 表格行 text = re.sub(r"\*\*([^*]*)\*\*", r"\1", text) # 粗体 text = re.sub(r"[|`]", "", text) # 竖线与反引号 text = re.sub(r"-{3,}", "", text) # 分隔线 text = re.sub(r"[ \t]+", " ", text) text = re.sub(r"\n{3,}", "\n\n", text) return text.strip("\n") def load_corpus(root: Path): """读取全部 .md,剥掉 frontmatter、代码块与 markdown 语法,返回(纯文本, 每文件字数)。""" chunks, per_file = [], Counter() for p in sorted(root.rglob("*.md")): text = p.read_text(encoding="utf-8") text = re.sub(r"^---\n.*?\n---\n", "", text, count=1, flags=re.S) text = re.sub(r"```.*?```", "", text, flags=re.S) # 围栏代码块 text = clean_text(text) if text: chunks.append(text) per_file[str(p.relative_to(root))] += len(text) return chunks, per_file def build_transitions(chunks): """二阶马尔可夫表:state=(前两字) -> 后继字符计数;starts=句子开头的大二元组。""" succ = defaultdict(Counter) starts = [] for text in chunks: if len(text) < 2: continue for i in range(len(text) - 2): succ[(text[i], text[i + 1])][text[i + 2]] += 1 for m in re.finditer(r"[。！？\n](.)(.)", text): a, b = m.group(1), m.group(2) if re.match(r"[\u4e00-\u9fffA-Za-z0-9]", a) and re.match(r"[\u4e00-\u9fffA-Za-z0-9]", b): starts.append((a, b)) return succ, starts def weighted_choice(counter, rng): r = rng.randrange(sum(counter.values())) for ch, c in counter.items(): r -= c if r < 0: return ch raise AssertionError("加权抽样落空 —— 计数表为空") def generate_dream(succ, starts, corpus_windows, rng, target, hard_cap): """随机游走拼一场梦;连读过长就强制跳转,保证是梦话不是背诵。""" out = list(rng.choice(starts) if starts else rng.choice(list(succ))) verbatim = 0 while len(out) < target or out[-1] not in "。！？": if len(out) >= hard_cap: break if verbatim > FORCE_JUMP_AFTER or rng.random() < JUMP_P: state = rng.choice(starts) if starts else rng.choice(list(succ)) out.extend(state) verbatim = 0 continue state = (out[-2], out[-1]) options = succ.get(state) if not options: state = rng.choice(starts) if starts else rng.choice(list(succ)) out.extend(state) verbatim = 0 continue nxt = weighted_choice(options, rng) out.append(nxt) tail = "".join(out[-WINDOW:]) verbatim = verbatim + 1 if len(tail) == WINDOW and tail in corpus_windows else 0 return "".join(out) def longest_verbatim(dream, corpus_windows): """梦与原文的最长连读(以 WINDOW 为最小刻度,>=WINDOW 的连读才计入)。""" best = 0 i = 0 while i <= len(dream) - WINDOW: if dream[i:i + WINDOW] in corpus_windows: j = i + WINDOW while j <= len(dream) - WINDOW and dream[j:j + WINDOW] in corpus_windows: j += WINDOW best = max(best, j - i + WINDOW - 1) i = j else: i += 1 return best def fragment_ratio(dream): """碎片度 = 梦里的不同大二元组 / 总大二元组;越高越碎,越像梦话。""" bigrams = [dream[i:i + 2] for i in range(len(dream) - 1)] return len(set(bigrams)) / len(bigrams) def murmur(succ, starts, corpus_windows, rng): """梦醒后残留的一句话:更短、更爱跳。""" return generate_dream(succ, starts, corpus_windows, rng, MURMUR_TARGET, MURMUR_TARGET * 3) def main(): seed = SEED count = DREAMS argv = sys.argv[1:] if "--seed" in argv: seed = int(argv[argv.index("--seed") + 1]) if "--count" in argv: count = int(argv[argv.index("--count") + 1]) rng = random.Random(seed) chunks, per_file = load_corpus(ROOT) succ, starts = build_transitions(chunks) vocab = set("".join(chunks)) corpus_windows = {"".join(chunks)[i:i + WINDOW] for i in range(len("".join(chunks)) - WINDOW + 1)} # ---- 断言查岗:先打脸再做梦 ---- assert len(chunks) >= 10, "语料页面太少,梦会变成复读机" assert len(vocab) > 500, "语料太贫瘠" assert len(succ) > 5000, "二阶转移表太小" assert starts, "一个句子开头都没找到?" print(f"=== 花园做梦机 · 种子 {seed} · {count} 场梦 ===") print(f"语料: {len(chunks)} 个页面 / {sum(len(c) for c in chunks):,} 字符 / " f"{len(vocab)} 个不同字 / {len(succ):,} 个二阶转移 / {len(starts):,} 个句子开头") print("梦的材料(按字数前八):") for path, n in per_file.most_common(8): print(f" {n:>6,} {path}") print() dreams = [] for k in range(1, count + 1): d = generate_dream(succ, starts, corpus_windows, rng, TARGET, HARD_CAP) m = murmur(succ, starts, corpus_windows, rng) frag = fragment_ratio(d) lv = longest_verbatim(d, corpus_windows) # ---- 梦的体检 ---- assert TARGET <= len(d) <= HARD_CAP, "梦的长度越界" assert d.endswith(("。", "！", "？")), "梦没做完就醒了" assert all(ch in vocab for ch in d), "梦里出现了语料之外的字" assert 0.4 <= frag <= 0.98, f"碎片度越界: {frag:.3f}" assert lv < MAX_VERBATIM, f"连读 {lv} 字,这是在背书不是在做梦" dreams.append((d, m, frag, lv)) print(f"--- 梦 {k} (长 {len(d)} · 碎片度 {frag:.3f} · 最长连读 {lv}) ---") print(d) print() print(f"梦醒残留: {m}") print() # ---- 梦与梦之间不许重复 ---- for i in range(count): for j in range(i + 1, count): assert dreams[i][0][:60] != dreams[j][0][:60], "两场梦撞了开头" print(f"全部断言通过。{count} 场梦,场场不同。") if __name__ == "__main__": main()