Tablinum-2: the Lite-Beam engram, release 2 (card, code, recipes, 9 released tables)
901f964 verified | """chunk.py IN.parquet OUT.parquet - finewiki enwiki articles -> engram passages (10-19). | |
| Each passage is ~WORDS words of one section, prefixed "Title — Section: " so it stands alone; each infobox becomes one | |
| "Title (infobox): key: value; ..." passage (dates, numbers, names - the facts low-bit weights lose first). Stubs under | |
| MIN_WORDS words and list/disambiguation pages are dropped. Output columns: pid (int64, file_idx<<32 | n), page_id, title, | |
| section, text.""" | |
| import sys, re, json | |
| import pyarrow as pa, pyarrow.parquet as pq | |
| WORDS, MIN_WORDS = 120, 30 | |
| SKIP = re.compile(r"^(List of|Lists of|Index of|Outline of)\b|\(disambiguation\)$") | |
| H = re.compile(r"^(#{1,6})\s+(.*)$") | |
| def passages(title, text, infob): | |
| out = [] | |
| try: | |
| for ib in json.loads(infob or "[]")[:2]: | |
| kv = "; ".join(f"{k}: {v}" for k, v in (ib.get("data") or {}).items() if v)[:900] | |
| if kv: | |
| out.append(("infobox", f"{title} (infobox): {kv}")) | |
| except (ValueError, AttributeError): | |
| pass | |
| sec, buf = "", [] | |
| def flush(): | |
| w = " ".join(buf).split() | |
| for i in range(0, len(w), WORDS): | |
| part = w[i:i + WORDS + 20] if len(w) - i <= WORDS + 20 else w[i:i + WORDS] | |
| if len(part) >= MIN_WORDS or (i == 0 and len(part) >= 12): | |
| out.append((sec, f"{title} — {sec}: " + " ".join(part) if sec else f"{title}: " + " ".join(part))) | |
| if len(w) - i <= WORDS + 20: | |
| break | |
| for line in text.split("\n"): | |
| m = H.match(line) | |
| if m: | |
| flush(); buf = [] | |
| sec = "" if len(m.group(1)) == 1 else m.group(2).strip() | |
| if sec.lower() in ("references", "external links", "see also", "further reading", "notes", "bibliography", "sources", "citations"): | |
| sec = "\x00" | |
| elif sec != "\x00" and line.strip(): | |
| buf.append(line.strip()) | |
| if sec != "\x00": | |
| flush() | |
| return out | |
| src, dst = sys.argv[1], sys.argv[2] | |
| fidx = int(re.search(r"(\d+)_(\d+)\.parquet$", src).group(2)) + 100 * int(re.search(r"(\d+)_(\d+)\.parquet$", src).group(1)) | |
| pf = pq.ParquetFile(src); w = None; n = 0; arts = 0 | |
| for rg in range(pf.num_row_groups): | |
| t = pf.read_row_group(rg, columns=["page_id", "title", "text", "infoboxes"]).to_pylist() | |
| cols = {"pid": [], "page_id": [], "title": [], "section": [], "text": []} | |
| for r in t: | |
| if SKIP.search(r["title"] or "") or len((r["text"] or "").split()) < MIN_WORDS: | |
| continue | |
| arts += 1 | |
| for sec, p in passages(r["title"], r["text"] or "", r["infoboxes"]): | |
| cols["pid"].append((fidx << 32) | n); cols["page_id"].append(r["page_id"]); cols["title"].append(r["title"]) | |
| cols["section"].append(sec); cols["text"].append(p); n += 1 | |
| tb = pa.table(cols) | |
| w = w or pq.ParquetWriter(dst + ".tmp", tb.schema, compression="zstd") | |
| w.write_table(tb, row_group_size=4096) # small row groups: the retriever reads single passages back by pid | |
| w.close() | |
| import os; os.replace(dst + ".tmp", dst) | |
| print(src, "articles", arts, "passages", n, flush=True) | |