| |
| """A/B the current extractor vs clean_v2 on retrieval quality, not aesthetics.""" |
| import sys, os, json, random, importlib.util |
| sys.path.insert(0, "/home/happy/.codex/bench/wiki_forensics") |
| os.environ.setdefault("TOKENIZERS_PARALLELISM", "false") |
| import libzim, numpy as np, clean_v2 |
| spec = importlib.util.spec_from_file_location("bp", "/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) |
| from transformers import AutoTokenizer |
| from sentence_transformers import SentenceTransformer |
| tok = AutoTokenizer.from_pretrained("intfloat/multilingual-e5-large") |
| model = SentenceTransformer("intfloat/multilingual-e5-large", device="cuda") |
|
|
| N = int(sys.argv[2]) if len(sys.argv) > 2 else 1200 |
| lang = sys.argv[1] |
| ZIM = {"ja":"wikipedia_ja_all_mini_2026-06.zim","zh":"wikipedia_zh_all_mini_2026-05.zim","en":"wikipedia_en_all_mini_2026-06.zim"}[lang] |
| ar = libzim.Archive("/home/happy/デスクトップ/mobius_ai/kiwix/"+ZIM) |
| step = max(1, ar.entry_count // (N*4)); arts = []; eid = 0 |
| while len(arts) < 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(): |
| raw = bytes(it.content).decode("utf-8","ignore") |
| if len(raw) > 2500: arts.append((e.title, raw)) |
| except Exception: pass |
| eid += step |
| print(f"{lang}: {len(arts)} articles") |
|
|
| def build(variant): |
| chunks, owner = [], [] |
| for i,(title, raw) in enumerate(arts): |
| text = bp.strip_html(raw) if variant == "current" else clean_v2.extract_text(raw) |
| cs = bp.chunk_text(text, tok) |
| if variant == "v2+title": |
| cs = [f"{title}\n{c}" for c in cs] |
| for c in cs: chunks.append(c); owner.append(i) |
| return chunks, np.array(owner) |
|
|
| res = {} |
| for variant in ("current", "v2", "v2+title"): |
| ch, own = build(variant) |
| ntok = sorted(len(x) for x in tok(["passage: "+c for c in ch], add_special_tokens=True)["input_ids"]) |
| emb = model.encode(["passage: "+c for c in ch], batch_size=256, normalize_embeddings=True, |
| convert_to_numpy=True, show_progress_bar=False).astype("float32") |
| q = model.encode([f"query: {t}" for t,_ in arts], batch_size=256, normalize_embeddings=True, |
| convert_to_numpy=True).astype("float32") |
| S = q @ emb.T |
| order = np.argsort(-S, axis=1)[:, :10] |
| hit1 = np.mean(own[order[:,0]] == np.arange(len(arts))) |
| rr = [] |
| for i in range(len(arts)): |
| r = np.where(own[order[i]] == i)[0] |
| rr.append(1.0/(r[0]+1) if len(r) else 0.0) |
| rnd = random.Random(3); idx = [rnd.randrange(len(ch)) for _ in range(4000)] |
| pair = float(np.mean(np.sum(emb[idx[:2000]] * emb[idx[2000:]], axis=1))) |
| foot = sum(1 for c in ch if "issued from Wikipedia" in c) / len(ch) |
| res[variant] = dict(chunks=len(ch), tok_p50=ntok[len(ntok)//2], tok_p95=ntok[len(ntok)*19//20], |
| over512=100*sum(1 for x in ntok if x>512)/len(ntok), |
| recall1=100*float(hit1), mrr=100*float(np.mean(rr)), |
| mean_pair_cos=round(pair,4), footer=round(100*foot,1)) |
| print(f" {variant:9s} chunks={res[variant]['chunks']:5d} tok p50={res[variant]['tok_p50']:4d} " |
| f">512={res[variant]['over512']:.2f}% recall@1={res[variant]['recall1']:5.1f}% MRR={res[variant]['mrr']:5.1f} " |
| f"pairwise-cos={res[variant]['mean_pair_cos']:.4f} footer={res[variant]['footer']}%", flush=True) |
| json.dump(res, open(f"/home/happy/.codex/bench/wiki_forensics/eval_{lang}.json","w"), indent=1) |
|
|
