calib-corpora / tools /gen_structured.py
worthant's picture
Restructure into a pool + per-model builds; add Muse-Glimmer-30B build
c9faf30 verified
Raw
History Blame Contribute Delete
4.38 kB
"""Structured data emitted inside a chat turn.
python tools/gen_structured.py --n 700
`pool/structured/repos.jsonl` holds real YAML, JSON, TOML and SQL as files.
That covers the tokens but not the situation the target model is actually in: an
agent asked for a config answers *inside* an assistant turn, so the structured
tokens sit next to chat special tokens rather than at the start of a document.
The structured payload in every conversation here is a real file from the pool,
byte for byte, keeping its upstream licence. Only the request and the sentence
introducing it are templated.
"""
from __future__ import annotations
import argparse
import os
import random
import sys
HERE = os.path.dirname(os.path.abspath(__file__))
sys.path.insert(0, HERE)
import poollib as P
SEED = 20260810
GENERATOR_VERSION = "gen_structured/1.0"
FENCE = {"yaml": "yaml", "json": "json", "toml": "toml", "sql": "sql"}
ASKS = [
("Show me the {lang} for `{path}` in {repo} — I want to copy it into a new project.",
"Here is `{path}` as it stands in {repo}:"),
("What does `{path}` configure in {repo}? Paste it and walk me through it.",
"`{path}` in {repo}, in full:"),
("I need the {lang} from `{path}`. Don't summarise it, I need the exact file.",
"Verbatim, `{path}`:"),
("Reproduce `{path}` from {repo} so I can diff it against mine.",
"`{path}` at the current revision:"),
]
TAILS = [
"The parts most projects need to change are the paths and the version pins; "
"everything else is defaults.",
"Note the ordering matters here — entries later in the file override earlier ones.",
"If you copy this, check the pinned versions first; they are the part that goes "
"stale fastest.",
"Nothing in here is project-specific except the names, so it ports cleanly.",
]
def main() -> int:
ap = argparse.ArgumentParser()
ap.add_argument("--n", type=int, default=700)
ap.add_argument("--seed", type=int, default=SEED)
ap.add_argument("--max-chars", type=int, default=9000,
help="skip payloads longer than this so one file cannot "
"swallow the whole category")
args = ap.parse_args()
rng = random.Random(args.seed)
src = [d for d in P.load_pool(categories=("structured",))
if d.get("render") == "text" and len(d["text"]) <= args.max_chars]
src.sort(key=lambda d: d["id"])
if not src:
raise SystemExit("no structured source documents in the pool")
print(f" {len(src)} structured source files eligible")
records = []
for i in range(min(args.n, len(src))):
d = src[(i * 4409) % len(src)]
lang = d["lang"]
ask, intro = ASKS[i % len(ASKS)]
fields = {"lang": lang.upper() if lang in ("sql", "json") else lang,
"path": d["path"], "repo": d["source"]}
body = f"```{FENCE.get(lang, lang)}\n{d['text']}\n```"
msgs = [
{"role": "user", "content": ask.format(**fields)},
{"role": "assistant",
"content": intro.format(**fields) + "\n\n" + body + "\n\n" +
TAILS[i % len(TAILS)]},
]
records.append(P.Record(
category="structured", domain="structured",
source=f"synthetic/structured-in-chat:{d['source']}",
license=d["license"], license_url=d.get("license_url", ""),
path=f"structured/chat/{d['source']}/{d['path']}", lang=lang,
origin="synth_structured_chat", messages=msgs,
reasoning_strength="low", render="chat", synthetic=True,
provenance={"generator": GENERATOR_VERSION, "seed": args.seed,
"payload_from": d["id"],
"payload_upstream": d.get("provenance", {}).get("upstream", ""),
"note": "the fenced payload is a real repository file, "
"reproduced byte for byte under its own licence; "
"only the request and framing sentences are templated"},
))
n = P.write_jsonl(os.path.join(P.POOL_ROOT, "structured", "synthetic",
"structured-in-chat.jsonl"), records)
print(f" pool/structured/synthetic/structured-in-chat.jsonl {n} conversations")
return 0
if __name__ == "__main__":
sys.exit(main())