| |
| """A/B the old fixed 1536-char chunker vs the new B+ sentence/token chunker on identical articles.""" |
| import sys, os, statistics, importlib.util |
| sys.path.insert(0, "/home/happy/デスクトップ/mobius_ai/MOBIUS_MMV") |
| os.environ["TOKENIZERS_PARALLELISM"] = "false" |
| import libzim |
| from transformers import AutoTokenizer |
| spec = importlib.util.spec_from_file_location("bplus", "/home/happy/デスクトップ/mobius_ai/MOBIUS_MMV/build_wiki_index_me5_bplus.py") |
| bp = importlib.util.module_from_spec(spec); spec.loader.exec_module(bp) |
| tok = AutoTokenizer.from_pretrained("intfloat/multilingual-e5-large") |
|
|
| OLD_CHARS, OLD_OVER = 1536, 256 |
| def old_chunk(t): |
| if not t or len(t) < 80: return [] |
| if len(t) <= OLD_CHARS: return [t] |
| out, s = [], 0 |
| while s < len(t): |
| c = t[s:s+OLD_CHARS].strip() |
| if c: out.append(c) |
| s += OLD_CHARS - OLD_OVER |
| return out |
|
|
| def stats(chunks, label): |
| if not chunks: return f"{label}: none" |
| n = [len(x) for x in tok(["passage: "+c for c in chunks], add_special_tokens=True)["input_ids"]] |
| n.sort(); ch = sorted(len(c) for c in chunks) |
| over = sum(1 for x in n if x > 512) |
| return (f"{label:9s} chunks={len(chunks):6d} chars p50={ch[len(ch)//2]:5d} p95={ch[len(ch)*19//20]:5d} max={ch[-1]:5d}" |
| f" tok p50={n[len(n)//2]:4d} p95={n[len(n)*19//20]:4d} max={n[-1]:5d} >512: {100*over/len(n):5.2f}%") |
|
|
| ZIM = {"zh":"/home/happy/デスクトップ/mobius_ai/kiwix/wikipedia_zh_all_mini_2026-05.zim", |
| "ja":"/home/happy/デスクトップ/mobius_ai/kiwix/wikipedia_ja_all_mini_2026-06.zim", |
| "en":"/home/happy/デスクトップ/mobius_ai/kiwix/wikipedia_en_all_mini_2026-03.zim"} |
| N = int(sys.argv[1]) if len(sys.argv) > 1 else 3000 |
| for lang, path in ZIM.items(): |
| if not os.path.exists(path): print(lang, "ZIM missing, skipped"); continue |
| ar = libzim.Archive(path); texts = [] |
| step = max(1, ar.entry_count // (N*3)) |
| eid = 0 |
| while len(texts) < N and eid < ar.entry_count: |
| try: |
| e = ar._get_entry_by_id(eid) |
| if not e.is_redirect: |
| it = e.get_item() |
| if "html" in str(it.mimetype).lower(): |
| t = bp.strip_html(bytes(it.content).decode("utf-8", "ignore")) |
| if len(t) >= 80: texts.append(t) |
| except Exception: pass |
| eid += step |
| old = [c for t in texts for c in old_chunk(t)] |
| new = [c for t in texts for c in bp.chunk_text(t, tok)] |
| print(f"\n=== {lang} articles={len(texts)} (mean article {statistics.mean(len(t) for t in texts):.0f} chars)") |
| print(" " + stats(old, "OLD 1536")) |
| print(" " + stats(new, "B+")) |
| print(f" chunks/article old={len(old)/len(texts):.2f} new={len(new)/len(texts):.2f}") |
| ex = max(new, key=len) |
| print(f" longest B+ chunk ends: …{ex[-70:]!r}") |
| print(f" sample B+ chunk: {new[len(new)//2][:150]!r}") |
|
|
