Datasets:
File size: 4,195 Bytes
2ff8533 | 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 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 | #!/usr/bin/env python3
"""Stage 2: download each needed upstream file ONE AT A TIME, pull out only the planned rows (low memory: row batches), convert with the
ORIGINAL build's functions (cyber/build_cyber_dataset.py) and append to gzip scratch files. Resumable (done-file list); deletes each
download right after use."""
import gzip, json, os, sys, time, collections
import pyarrow.parquet as pq
import pandas as pd
import huggingface_hub as hf
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
sys.path.insert(0, os.path.join(os.path.dirname(os.path.abspath(__file__)), ".."))
from common import *
import build_cyber_dataset as orig # original converters: messages_from_*, cve_answer, CVE_SYSTEM, to_native
plan = json.load(open(os.path.join(SCRATCH, "plan.json")))
done_path = os.path.join(SCRATCH, "extract_done.json")
done = set(json.load(open(done_path))) if os.path.exists(done_path) else set()
t0 = time.time()
def drop(path):
for p in {path, os.path.realpath(path)}:
try: os.remove(p)
except OSError: pass
def get_out(outs, stream):
# NOT outs.setdefault(stream, writer(stream)): that opens a NEW gzip handle for every row and corrupts the file
if stream not in outs:
outs[stream] = writer(stream)
return outs[stream]
def writer(stream):
return gzip.open(os.path.join(SCRATCH, "rows_" + stream.replace("/", "_").replace(":", "_") + ".jsonl.gz"), "at", encoding="utf-8", compresslevel=3)
# group candidates: (repo, file) -> {id: (stream, rank)}
work = collections.defaultdict(dict)
for stream, s in plan["streams"].items():
repo = SRC_NAME["swehero"] if stream == "swehero" else SRC_NAME["openswe"] if stream.startswith("openswe") else SRC_NAME["cve"]
for rank, rid, f in s["candidates"]:
work[(repo, f)][rid] = (stream, rank)
print(f"{len(work)} upstream files to process, {sum(len(v) for v in work.values())} candidate rows", flush=True)
for n, ((repo, f), need) in enumerate(sorted(work.items()), 1):
key = repo + "|" + f
if key in done:
continue
kw = {"revision": "refs/convert/parquet"} if repo == SRC_NAME["cve"] else {}
path = hf.hf_hub_download(repo, f, repo_type="dataset", **kw)
got = 0
outs = {}
if repo == SRC_NAME["cve"]:
df = pd.read_parquet(path).drop_duplicates("CVE-ID").set_index("CVE-ID", drop=False)
for rid, (stream, rank) in sorted(need.items(), key=lambda x: x[1][1]):
cid = rid[len("cve-"):]
if cid not in df.index: continue
r = df.loc[cid]
ans = orig.cve_answer(r)
if len(ans) < 20: continue
msgs = [{"role": "system", "content": orig.CVE_SYSTEM}, {"role": "user", "content": f"What is {cid} and how severe is it?"}, {"role": "assistant", "content": ans}]
get_out(outs, stream).write(json.dumps(orig.to_native({"rk": rank, "id": rid, "source": repo, "messages": msgs}), ensure_ascii=False) + "\n")
got += 1
else:
sw = repo == SRC_NAME["swehero"]
cols = ["trajectory_id", "trajectory"] if sw else ["trajectory_id", "messages", "resolved"]
pf = pq.ParquetFile(path)
for batch in pf.iter_batches(batch_size=50, columns=cols):
d = batch.to_pydict()
for i, tid in enumerate(d["trajectory_id"]):
rid = ("swehero-" if sw else "openswe-") + str(tid)
if rid not in need: continue
if not sw and d["resolved"][i] != 1: continue
msgs = (orig.messages_from_swehero_trajectory if sw else orig.messages_from_openswe)(list(d["trajectory" if sw else "messages"][i]))
if len(msgs) < 2: continue
stream, rank = need[rid]
get_out(outs, stream).write(json.dumps(orig.to_native({"rk": rank, "id": rid, "source": repo, "messages": msgs}), ensure_ascii=False) + "\n")
got += 1
for o in outs.values(): o.close()
drop(path)
done.add(key); json.dump(sorted(done), open(done_path, "w"))
print(f"[{n}/{len(work)}] {repo.split('/')[-1][:28]:28s} {f[-52:]:52s} kept {got}/{len(need)} {time.time()-t0:.0f}s", flush=True)
print("EXTRACT_DONE", flush=True)
|