File size: 3,396 Bytes
901f964 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 | """dolma_chunk.py MODE IN.json.gz OUT.parquet - Common Pile (dolma json.gz: id, text, metadata) -> engram passages (10-20,
PLAN-verticals B2/B4-6). Same schema as chunk.py; file_idx from OUT's name AAA_BB.
doc: OER books and pages (libretexts, pressbooks, oercommons, doab). Title = metadata.title; a short line with no end punctuation starts a new section; WORDS-word windows.
cap: Caselaw Access Project opinions. Title "Case name (court, date)" from the caption lines; the caption/counsel block is
skipped (body starts at the first line of >= 25 words); only the first CAP_WINDOWS windows are kept (the syllabus /
statement of the holding), and opinions with a body under 120 words are dropped - 6.7M cases would otherwise be ~140M
passages."""
import sys, os, re, json, gzip
import pyarrow as pa, pyarrow.parquet as pq
WORDS, MIN_WORDS, CAP_WINDOWS = 140, 12, int(os.environ.get("CAP_WINDOWS", 3))
mode, src, dst = sys.argv[1:4]
a, b = re.search(r"(\d+)_(\d+)\.parquet$", dst).groups(); fidx = int(b) + 100 * int(a)
cols = {"pid": [], "page_id": [], "title": [], "section": [], "text": []}; n = docs = 0
def emit(title, sec, words, doc, limit=None):
global n
k = 0
for i in range(0, len(words), WORDS):
part = words[i:i + WORDS + 20] if len(words) - i <= WORDS + 20 else words[i:i + WORDS]
if len(part) >= MIN_WORDS:
cols["pid"].append((fidx << 32) | n); cols["page_id"].append(doc); cols["title"].append(title); cols["section"].append(sec)
cols["text"].append(f"{title} — {sec}: " + " ".join(part) if sec else f"{title}: " + " ".join(part)); n += 1; k += 1
if len(words) - i <= WORDS + 20 or (limit and k >= limit):
break
for line in gzip.open(src, "rt"):
d = json.loads(line); md = d.get("metadata") or {}; t = d["text"]; docs += 1
if mode == "cap":
lines = [x.strip() for x in t.split("\n") if x.strip()]
if not lines:
continue
start = next((i for i, x in enumerate(lines) if len(x.split()) >= 25), None)
if start is None:
continue
body = " ".join(lines[start:]).split()
if len(body) < 120:
continue
cap = [x for x in lines[:start] if not x.lower().startswith(("no.", "docket", "before", "argued", "submitted"))]
court = next((x for x in cap[1:6] if re.search(r"court|supreme|appeals|circuit", x, re.I)), "")
date = next((x for x in cap[1:8] if re.search(r"\b1[6-9]\d\d\b|\b20\d\d\b", x) and len(x) < 40), "")
title = re.sub(r"\s+", " ", cap[0] if cap else lines[0])[:160] + (f" ({', '.join(x for x in (court, date) if x)})" if court or date else "")
emit(title, "", body, docs, CAP_WINDOWS)
else:
title = (md.get("title") or t.split("\n", 1)[0])[:160].strip()
sec, buf = "", []
for x in t.split("\n"):
x = x.strip().strip("|").strip()
if not x:
continue
if len(x.split()) <= 8 and not re.search(r"[.:;,?!)\]]$", x) and len(buf) >= MIN_WORDS:
emit(title, sec, buf, docs); buf = []; sec = x[:100]
else:
buf += x.split()
emit(title, sec, buf, docs)
pq.write_table(pa.table(cols), dst + ".tmp", compression="zstd", row_group_size=4096); os.replace(dst + ".tmp", dst)
print(src, mode, "docs", docs, "passages", n, flush=True)
|