Datasets:

AtomicChat
/

calib-corpora

"""Build a per-model calibration set from the pool under a recipe.
    python tools/build.py --recipe recipes/muse-glimmer-30b.yaml
Deterministic: same pool, same recipe, same seed, same bytes out.  Ordering
never depends on dictionary iteration or on the filesystem — documents are
selected by a seeded shuffle of an id-sorted list, and the output is written in
that order.
Dialogue records are rendered here, not in the pool, so the special-token
sequences in the output are this model's and nobody else's.  Records marked
`render: dsv4` or carrying `provenance.excluded_from_builds` are skipped with a
reason; they stay in the pool as history.
"""

from __future__ import annotations

import argparse
import hashlib
import json
import os
import random
import subprocess
import sys
from collections import defaultdict

HERE = os.path.dirname(os.path.abspath(__file__))
sys.path.insert(0, HERE)

import glimmer_fmt as G          # default renderer; see FORMATTERS below

# chat.format in the recipe -> renderer module. A recipe with no chat.format
# builds exactly as it did before this was added.
FORMATTERS = {
    "auto":     "auto_fmt",
    "glimmer":  "glimmer_fmt",
    "nemotron": "nemotron_fmt",
}
import poollib as P

BUILD_VERSION = "build.py/1.0"
DOC_SEP = "\n\n"


def load_recipe(path: str) -> dict:
    try:
        import yaml
        with open(path, encoding="utf-8") as f:
            return yaml.safe_load(f)
    except ImportError:
        raise SystemExit("PyYAML is required to read recipes")


def preflight(recipe: dict, repo: str) -> None:
    """Everything that can be wrong before a single token is read.
    Each check names the file it wanted, why, and the command that fixes it.
    A traceback from deep inside a tokenizer says none of that.
    """
    problems = []

    name = recipe.get("name")
    if not name:
        problems.append("the recipe has no `name`. It sets the build directory "
                        "and selects the vocabulary sweep, so it is required.")

    chat = recipe.get("chat", {})
    fmt = chat.get("format", "glimmer")
    if fmt not in FORMATTERS:
        problems.append(
            "chat.format is %r, which is not a renderer.\n"
            "    known renderers: %s" % (fmt, ", ".join(sorted(FORMATTERS))))
    if fmt == "auto":
        md = os.environ.get("FOUNDRY_MODEL_DIR", "/src")
        if not os.path.isdir(md):
            problems.append(
                "chat.format is `auto`, which renders through the model's own\n"
                "    chat template, but there are no weights at %s.\n"
                "    fix: export FOUNDRY_MODEL_DIR=/path/to/original/weights\n"
                "         or download them:  get_upstream" % md)

    tok_path = (recipe.get("model") or {}).get("tokenizer")
    if not tok_path:
        problems.append("model.tokenizer is not set in the recipe.")
    elif not os.path.isfile(tok_path):
        problems.append(
            "no tokenizer at %s\n"
            "    Token budgets cannot be counted without it.\n"
            "    fix: download the original weights (get_upstream), or point\n"
            "         model.tokenizer at where they already are." % tok_path)

    pool_dir = os.path.join(repo, "pool")
    if not os.path.isdir(pool_dir):
        problems.append("no pool at %s" % pool_dir)
    elif not any(f.endswith(".jsonl")
                 for _, _, fs in os.walk(pool_dir) for f in fs):
        problems.append(
            "the pool at %s holds no .jsonl files.\n"
            "    A clone without git-lfs brings down pointer stubs.\n"
            "    fix: git lfs pull, or run check_pool then fix_pool" % pool_dir)

    if problems:
        print("cannot build %s\n" % (recipe.get("name") or "this recipe"),
              file=sys.stderr)
        for i, p in enumerate(problems, 1):
            print("  %d. %s\n" % (i, p), file=sys.stderr)
        raise SystemExit(1)

    sweep = os.path.join(repo, "pool", "vocab_sweep", "synthetic",
                         "%s.jsonl" % name)
    if not os.path.isfile(sweep):
        print("WARNING: no vocabulary sweep at %s" % sweep)
        print("         The vocab_sweep share will come out empty and the")
        print("         corpus will silently lose it. Generate it first:")
        print("           python tools/vocab_sweep.py --tokenizer %s --name %s"
              % (tok_path, name))
        print()


def usable(rec: dict) -> tuple[bool, str]:
    if rec.get("render") == "dsv4":
        return False, "render=dsv4 (DeepSeek markup, not portable)"
    why = (rec.get("provenance") or {}).get("excluded_from_builds")
    if why:
        return False, why
    return True, ""


def make_renderer(recipe: dict):
    chat = recipe.get("chat", {})

    fmt_name = chat.get("format", "glimmer")
    if fmt_name not in FORMATTERS:
        raise SystemExit("unknown chat.format %r; known: %s"
                         % (fmt_name, ", ".join(sorted(FORMATTERS))))
    import importlib
    F = importlib.import_module(FORMATTERS[fmt_name])

    # chat.reasoning_effort: true (auto renderer only) passes each document's effort to the template
    extra = {"reasoning_effort": True} if fmt_name == "auto" and chat.get("reasoning_effort") else {}

    def render(rec: dict) -> str:
        prov = rec.get("provenance") or {}
        return F.render(
            rec["messages"],
            tools=rec.get("tools"),
            reasoning_strength=rec.get("reasoning_strength") or F.DEFAULT_REASONING,
            knowledge_cutoff=chat.get("knowledge_cutoff", F.DEFAULT_KNOWLEDGE_CUTOFF),
            current_date=chat.get("current_date"),
            namespace_descriptions=prov.get("namespace_descriptions"),
            add_generation_prompt=False,
            add_bos=chat.get("add_bos_per_document", True),
            **extra,
        )
    return render


def seeded_order(recs: list[dict], seed: int) -> list[dict]:
    """Shuffle reproducibly: sort by id, then permute with a seeded RNG."""
    out = sorted(recs, key=lambda r: r["id"])
    random.Random(seed).shuffle(out)
    return out


def truncate_to_tokens(text: str, tok: P.TargetTokenizer, lo: int, hi: int) -> str | None:
    """A contiguous prefix of `text` between lo and hi tokens, cut on a line
    boundary so the document stays unbroken prose/code rather than ending
    mid-token."""
    n = tok.count(text)
    if n < lo:
        return None
    if n <= hi:
        return text
    lines = text.split("\n")
    # one batched encode rather than a tokenizer call per line; the per-line
    # counts only have to be close enough to find the cut, and the result is
    # re-counted exactly below
    per_line = [len(ids) + 1 for ids in tok.encode_batch(lines)]
    keep, acc = [], 0
    for ln, c in zip(lines, per_line):
        if acc + c > hi:
            break
        keep.append(ln)
        acc += c
    body = "\n".join(keep)
    n = tok.count(body)
    while n > hi and keep:
        keep = keep[:int(len(keep) * 0.95) or len(keep) - 1]
        body = "\n".join(keep)
        n = tok.count(body)
    return body if n >= lo else None


def main() -> int:
    ap = argparse.ArgumentParser()
    ap.add_argument("--recipe", required=True)
    ap.add_argument("--out", default=None)
    ap.add_argument("--repo", default=P.REPO_ROOT)
    args = ap.parse_args()

    recipe = load_recipe(args.recipe)
    preflight(recipe, args.repo)
    name = recipe["name"]
    seed = int(recipe.get("seed", 0))
    model = recipe["model"]
    out_dir = args.out or os.path.join(args.repo, "builds", name)
    os.makedirs(out_dir, exist_ok=True)

    tok = P.TargetTokenizer(model["tokenizer"], n_vocab=model.get("vocab_size"))
    render_chat = make_renderer(recipe)
    print(f"recipe {name}  seed {seed}  tokenizer {model['tokenizer']}")

    docs = P.load_pool()
    docs = [d for d in docs if not d["_shard"].startswith(os.path.join("pool", "_quarantine"))]
    # The vocabulary sweep is a function of the tokenizer, so it is sharded one
    # file per target model: pool/vocab_sweep/synthetic/<recipe name>.jsonl.
    # auto-select ours and drop every other model's sweep, so adding a model
    # never requires editing the recipes that already exist.
    SWEEP_DIR = "pool/vocab_sweep/synthetic/"
    mine = SWEEP_DIR + name + ".jsonl"
    before = len(docs)
    docs = [d for d in docs
            if not (lambda s: s.startswith(SWEEP_DIR) and s != mine)(d["_shard"].replace(os.sep, "/"))]
    dropped = before - len(docs)
    if dropped:
        print(f"vocab sweep: using {mine}, dropped {dropped:,} documents from other models' sweeps")

    # escape hatch for anything else that is model-specific
    shard_deny = (recipe.get("pool") or {}).get("shard_deny") or []
    if shard_deny:
        before = len(docs)
        docs = [d for d in docs if not any(s in d["_shard"].replace(os.sep, "/") for s in shard_deny)]
        print(f"shard_deny: dropped {before - len(docs):,} documents from {len(shard_deny)} pattern(s)")
    skipped: dict[str, int] = defaultdict(int)
    pool: dict[str, list[dict]] = defaultdict(list)
    for d in docs:
        ok, why = usable(d)
        if not ok:
            skipped[why] += 1
            continue
        pool[d["category"]].append(d)
    print(f"pool: {sum(len(v) for v in pool.values()):,} usable documents")
    for why, n in sorted(skipped.items(), key=lambda kv: -kv[1]):
        print(f"  skipped {n:5}  {why[:96]}")

    # ------------------------------------------------------------ calib_train
    spec = recipe["calib_train"]
    target = int(spec["target_tokens"])
    shares = spec["shares"]
    total_share = sum(shares.values())
    # No single document may dominate its slice. An amalgamated single-header or
    # a generated search index is one file, and left uncapped one of them takes
    # 30-60% of a category -- the imatrix then describes that file rather than
    # the category. Overridable per category because a longctx document is
    # supposed to be large.
    frac = dict(spec.get("max_doc_fraction") or {})
    default_frac = float(frac.pop("default", 0.05))
    chosen: list[tuple[dict, str, int]] = []
    actual: dict[str, int] = {}
    oversize: list[tuple[str, str, str, int]] = []
    print(f"\ncalib_train: target {target:,} tokens")
    for cat in sorted(shares):
        budget = int(target * shares[cat] / total_share)
        cap = int(budget * float(frac.get(cat, default_frac)))
        cands = seeded_order(pool.get(cat, []), seed + sum(map(ord, cat)))
        got, taken, skipped_big = 0, 0, 0
        for d in cands:
            if got >= budget:
                break
            text = render_chat(d) if d.get("render") == "chat" else d["text"]
            n = tok.count(text)
            if n == 0:
                continue
            if n > cap:
                skipped_big += 1
                oversize.append((cat, d["source"], d["path"], n))
                continue
            chosen.append((d, text, n))
            got += n
            taken += 1
        actual[cat] = got
        avail = "" if got >= budget * 0.98 else "  <-- pool exhausted"
        big = f"  ({skipped_big} over the {cap:,}-token cap)" if skipped_big else ""
        print(f"  {cat:13} target {budget:9,}  got {got:9,} in {taken:5} docs{avail}{big}")
    if oversize:
        print(f"\n  {len(oversize)} documents skipped as too large for their slice:")
        for cat, src, path, n in sorted(oversize, key=lambda t: -t[3])[:8]:
            print(f"    {cat:12} {n:8,} tok  {src[:16]:16} {path[:52]}")

    chosen = sorted(chosen, key=lambda t: (t[0]["category"], t[0]["id"]))
    random.Random(seed).shuffle(chosen)
    total_tokens = sum(n for _, _, n in chosen)
    body = DOC_SEP.join(t for _, t, _ in chosen) + "\n"
    train_path = os.path.join(out_dir, "calib_train.txt")
    with open(train_path, "w", encoding="utf-8") as f:
        f.write(body)
    print(f"  wrote {train_path}  {len(chosen):,} docs  {total_tokens:,} tokens "
          f"({len(body):,} chars)")

    with open(os.path.join(out_dir, "calib_train.manifest.jsonl"), "w",
              encoding="utf-8") as f:
        for d, text, n in chosen:
            f.write(json.dumps({
                "id": d["id"], "category": d["category"], "source": d["source"],
                "license": d["license"], "path": d["path"], "lang": d["lang"],
                "origin": d["origin"], "synthetic": d.get("synthetic", False),
                "render": d.get("render", "text"),
                "reasoning_strength": d.get("reasoning_strength", ""),
                "tokens": n, "chars": len(text), "shard": d["_shard"],
            }, ensure_ascii=False) + "\n")

    # ---------------------------------------------------------- calib_longctx
    lspec = recipe.get("calib_longctx")
    long_docs: list[tuple[dict, str, int]] = []
    if lspec:
        lo, hi = int(lspec["doc_tokens_min"]), int(lspec["doc_tokens_max"])
        ltarget = int(lspec["target_tokens"])
        used = {d["id"] for d, _, _ in chosen}
        cands = seeded_order(
            [d for c in lspec.get("sources", ["longctx"]) for d in pool.get(c, [])
             if d["id"] not in used], seed + 977)
        got = 0
        print(f"\ncalib_longctx: target {ltarget:,} tokens, "
              f"documents of {lo:,}-{hi:,} tokens, unbroken")
        for d in cands:
            if got >= ltarget:
                break
            text = render_chat(d) if d.get("render") == "chat" else d["text"]
            cut = truncate_to_tokens(text, tok, lo, hi)
            if cut is None:
                continue
            n = tok.count(cut)
            long_docs.append((d, cut, n))
            got += n
        lpath = os.path.join(out_dir, "calib_longctx.txt")
        with open(lpath, "w", encoding="utf-8") as f:
            f.write(DOC_SEP.join(t for _, t, _ in long_docs) + "\n")
        lens = sorted(n for _, _, n in long_docs)
        print(f"  wrote {lpath}  {len(long_docs)} docs  {got:,} tokens  "
              f"min {lens[0]:,} max {lens[-1]:,}" if lens else "  no documents qualified")
        with open(os.path.join(out_dir, "calib_longctx.manifest.jsonl"), "w",
                  encoding="utf-8") as f:
            for d, text, n in long_docs:
                f.write(json.dumps({
                    "id": d["id"], "category": d["category"], "source": d["source"],
                    "license": d["license"], "path": d["path"], "origin": d["origin"],
                    "tokens": n, "chars": len(text), "shard": d["_shard"],
                }, ensure_ascii=False) + "\n")

    # ----------------------------------------------------------------- manifest
    shard_hashes = {}
    for p in P.shard_paths(P.POOL_ROOT):
        shard_hashes[os.path.relpath(p, args.repo)] = {
            "sha256": P.sha256_file(p),
            "bytes": os.path.getsize(p),
            "documents": sum(1 for _ in P.read_jsonl(p)),
        }

    lens = sorted(n for _, _, n in chosen)

    def pct(q: float) -> int:
        return lens[min(len(lens) - 1, int(q * len(lens)))] if lens else 0

    by_cat_docs = defaultdict(int)
    by_cat_tok = defaultdict(int)
    synth_tok = 0
    for d, _, n in chosen:
        by_cat_docs[d["category"]] += 1
        by_cat_tok[d["category"]] += n
        if d.get("synthetic"):
            synth_tok += n

    manifest = {
        "build": name,
        "built_at": recipe.get("chat", {}).get("current_date"),
        "build_tool_version": BUILD_VERSION,
        "recipe": {"path": os.path.relpath(args.recipe, args.repo),
                   "sha256": P.sha256_file(args.recipe), "seed": seed},
        "model": model,
        "tokenizer": {
            "path": model["tokenizer"],
            "sha256": P.sha256_file(model["tokenizer"]),
            "rows": tok.n_vocab,
            "pre_tokenizer": model.get("pre_tokenizer", "llama4"),
        },
        "llama_cpp": {
            "commit": model.get("llama_cpp_commit", ""),
            "note": "llama-imatrix must be given --parse-special or the chat "
                    "markup in calib_train.txt is tokenised as literal text "
                    "(parse_special defaults to false in that tool)",
        },
        "calib_train": {
            "path": "calib_train.txt",
            "sha256": P.sha256_file(train_path),
            "bytes": os.path.getsize(train_path),
            "documents": len(chosen),
            "tokens": total_tokens,
            "target_tokens": target,
            "document_separator": DOC_SEP,
            "shares_requested_percent": {k: round(100.0 * v / total_share, 3)
                                         for k, v in sorted(shares.items())},
            "shares_actual_percent": {k: round(100.0 * by_cat_tok[k] / total_tokens, 3)
                                      for k in sorted(by_cat_tok)},
            "documents_by_category": dict(sorted(by_cat_docs.items())),
            "tokens_by_category": dict(sorted(by_cat_tok.items())),
            "synthetic_tokens": synth_tok,
            "synthetic_percent": round(100.0 * synth_tok / total_tokens, 3),
            "document_tokens": {
                "p50": pct(0.50), "p90": pct(0.90), "p95": pct(0.95),
                "p99": pct(0.99), "max": lens[-1] if lens else 0,
                "min": lens[0] if lens else 0,
            },
        },
        "pool_sources": dict(sorted(shard_hashes.items())),
    }
    if long_docs:
        llens = sorted(n for _, _, n in long_docs)
        manifest["calib_longctx"] = {
            "path": "calib_longctx.txt",
            "sha256": P.sha256_file(os.path.join(out_dir, "calib_longctx.txt")),
            "documents": len(long_docs),
            "tokens": sum(llens),
            "document_tokens": {"min": llens[0], "p50": llens[len(llens) // 2],
                                "max": llens[-1]},
            "contiguous": True,
        }
    with open(os.path.join(out_dir, "manifest.json"), "w", encoding="utf-8") as f:
        json.dump(manifest, f, indent=2, ensure_ascii=False)
        f.write("\n")
    print(f"\nwrote {os.path.join(out_dir, 'manifest.json')}")
    print(json.dumps(manifest["calib_train"]["shares_actual_percent"], indent=2))
    return 0


if __name__ == "__main__":
    sys.exit(main())

	"""Build a per-model calibration set from the pool under a recipe.

	python tools/build.py --recipe recipes/muse-glimmer-30b.yaml

	Deterministic: same pool, same recipe, same seed, same bytes out. Ordering
	never depends on dictionary iteration or on the filesystem — documents are
	selected by a seeded shuffle of an id-sorted list, and the output is written in
	that order.

	Dialogue records are rendered here, not in the pool, so the special-token
	sequences in the output are this model's and nobody else's. Records marked
	`render: dsv4` or carrying `provenance.excluded_from_builds` are skipped with a
	reason; they stay in the pool as history.
	"""

	from __future__ import annotations

	import argparse
	import hashlib
	import json
	import os
	import random
	import subprocess
	import sys
	from collections import defaultdict

	HERE = os.path.dirname(os.path.abspath(__file__))
	sys.path.insert(0, HERE)

	import glimmer_fmt as G # default renderer; see FORMATTERS below

	# chat.format in the recipe -> renderer module. A recipe with no chat.format
	# builds exactly as it did before this was added.
	FORMATTERS = {
	"auto": "auto_fmt",
	"glimmer": "glimmer_fmt",
	"nemotron": "nemotron_fmt",
	}
	import poollib as P

	BUILD_VERSION = "build.py/1.0"
	DOC_SEP = "\n\n"


	def load_recipe(path: str) -> dict:
	try:
	import yaml
	with open(path, encoding="utf-8") as f:
	return yaml.safe_load(f)
	except ImportError:
	raise SystemExit("PyYAML is required to read recipes")


	def preflight(recipe: dict, repo: str) -> None:
	"""Everything that can be wrong before a single token is read.

	Each check names the file it wanted, why, and the command that fixes it.
	A traceback from deep inside a tokenizer says none of that.
	"""
	problems = []

	name = recipe.get("name")
	if not name:
	problems.append("the recipe has no `name`. It sets the build directory "
	"and selects the vocabulary sweep, so it is required.")

	chat = recipe.get("chat", {})
	fmt = chat.get("format", "glimmer")
	if fmt not in FORMATTERS:
	problems.append(
	"chat.format is %r, which is not a renderer.\n"
	" known renderers: %s" % (fmt, ", ".join(sorted(FORMATTERS))))
	if fmt == "auto":
	md = os.environ.get("FOUNDRY_MODEL_DIR", "/src")
	if not os.path.isdir(md):
	problems.append(
	"chat.format is `auto`, which renders through the model's own\n"
	" chat template, but there are no weights at %s.\n"
	" fix: export FOUNDRY_MODEL_DIR=/path/to/original/weights\n"
	" or download them: get_upstream" % md)

	tok_path = (recipe.get("model") or {}).get("tokenizer")
	if not tok_path:
	problems.append("model.tokenizer is not set in the recipe.")
	elif not os.path.isfile(tok_path):
	problems.append(
	"no tokenizer at %s\n"
	" Token budgets cannot be counted without it.\n"
	" fix: download the original weights (get_upstream), or point\n"
	" model.tokenizer at where they already are." % tok_path)

	pool_dir = os.path.join(repo, "pool")
	if not os.path.isdir(pool_dir):
	problems.append("no pool at %s" % pool_dir)
	elif not any(f.endswith(".jsonl")
	for _, _, fs in os.walk(pool_dir) for f in fs):
	problems.append(
	"the pool at %s holds no .jsonl files.\n"
	" A clone without git-lfs brings down pointer stubs.\n"
	" fix: git lfs pull, or run check_pool then fix_pool" % pool_dir)

	if problems:
	print("cannot build %s\n" % (recipe.get("name") or "this recipe"),
	file=sys.stderr)
	for i, p in enumerate(problems, 1):
	print(" %d. %s\n" % (i, p), file=sys.stderr)
	raise SystemExit(1)

	sweep = os.path.join(repo, "pool", "vocab_sweep", "synthetic",
	"%s.jsonl" % name)
	if not os.path.isfile(sweep):
	print("WARNING: no vocabulary sweep at %s" % sweep)
	print(" The vocab_sweep share will come out empty and the")
	print(" corpus will silently lose it. Generate it first:")
	print(" python tools/vocab_sweep.py --tokenizer %s --name %s"
	% (tok_path, name))
	print()


	def usable(rec: dict) -> tuple[bool, str]:
	if rec.get("render") == "dsv4":
	return False, "render=dsv4 (DeepSeek markup, not portable)"
	why = (rec.get("provenance") or {}).get("excluded_from_builds")
	if why:
	return False, why
	return True, ""


	def make_renderer(recipe: dict):
	chat = recipe.get("chat", {})

	fmt_name = chat.get("format", "glimmer")
	if fmt_name not in FORMATTERS:
	raise SystemExit("unknown chat.format %r; known: %s"
	% (fmt_name, ", ".join(sorted(FORMATTERS))))
	import importlib
	F = importlib.import_module(FORMATTERS[fmt_name])

	# chat.reasoning_effort: true (auto renderer only) passes each document's effort to the template
	extra = {"reasoning_effort": True} if fmt_name == "auto" and chat.get("reasoning_effort") else {}

	def render(rec: dict) -> str:
	prov = rec.get("provenance") or {}
	return F.render(
	rec["messages"],
	tools=rec.get("tools"),
	reasoning_strength=rec.get("reasoning_strength") or F.DEFAULT_REASONING,
	knowledge_cutoff=chat.get("knowledge_cutoff", F.DEFAULT_KNOWLEDGE_CUTOFF),
	current_date=chat.get("current_date"),
	namespace_descriptions=prov.get("namespace_descriptions"),
	add_generation_prompt=False,
	add_bos=chat.get("add_bos_per_document", True),
	**extra,
	)
	return render


	def seeded_order(recs: list[dict], seed: int) -> list[dict]:
	"""Shuffle reproducibly: sort by id, then permute with a seeded RNG."""
	out = sorted(recs, key=lambda r: r["id"])
	random.Random(seed).shuffle(out)
	return out


	def truncate_to_tokens(text: str, tok: P.TargetTokenizer, lo: int, hi: int) -> str \| None:
	"""A contiguous prefix of `text` between lo and hi tokens, cut on a line
	boundary so the document stays unbroken prose/code rather than ending
	mid-token."""
	n = tok.count(text)
	if n < lo:
	return None
	if n <= hi:
	return text
	lines = text.split("\n")
	# one batched encode rather than a tokenizer call per line; the per-line
	# counts only have to be close enough to find the cut, and the result is
	# re-counted exactly below
	per_line = [len(ids) + 1 for ids in tok.encode_batch(lines)]
	keep, acc = [], 0
	for ln, c in zip(lines, per_line):
	if acc + c > hi:
	break
	keep.append(ln)
	acc += c
	body = "\n".join(keep)
	n = tok.count(body)
	while n > hi and keep:
	keep = keep[:int(len(keep) * 0.95) or len(keep) - 1]
	body = "\n".join(keep)
	n = tok.count(body)
	return body if n >= lo else None


	def main() -> int:
	ap = argparse.ArgumentParser()
	ap.add_argument("--recipe", required=True)
	ap.add_argument("--out", default=None)
	ap.add_argument("--repo", default=P.REPO_ROOT)
	args = ap.parse_args()

	recipe = load_recipe(args.recipe)
	preflight(recipe, args.repo)
	name = recipe["name"]
	seed = int(recipe.get("seed", 0))
	model = recipe["model"]
	out_dir = args.out or os.path.join(args.repo, "builds", name)
	os.makedirs(out_dir, exist_ok=True)

	tok = P.TargetTokenizer(model["tokenizer"], n_vocab=model.get("vocab_size"))
	render_chat = make_renderer(recipe)
	print(f"recipe {name} seed {seed} tokenizer {model['tokenizer']}")

	docs = P.load_pool()
	docs = [d for d in docs if not d["_shard"].startswith(os.path.join("pool", "_quarantine"))]
	# The vocabulary sweep is a function of the tokenizer, so it is sharded one
	# file per target model: pool/vocab_sweep/synthetic/<recipe name>.jsonl.
	# auto-select ours and drop every other model's sweep, so adding a model
	# never requires editing the recipes that already exist.
	SWEEP_DIR = "pool/vocab_sweep/synthetic/"
	mine = SWEEP_DIR + name + ".jsonl"
	before = len(docs)
	docs = [d for d in docs
	if not (lambda s: s.startswith(SWEEP_DIR) and s != mine)(d["_shard"].replace(os.sep, "/"))]
	dropped = before - len(docs)
	if dropped:
	print(f"vocab sweep: using {mine}, dropped {dropped:,} documents from other models' sweeps")

	# escape hatch for anything else that is model-specific
	shard_deny = (recipe.get("pool") or {}).get("shard_deny") or []
	if shard_deny:
	before = len(docs)
	docs = [d for d in docs if not any(s in d["_shard"].replace(os.sep, "/") for s in shard_deny)]
	print(f"shard_deny: dropped {before - len(docs):,} documents from {len(shard_deny)} pattern(s)")
	skipped: dict[str, int] = defaultdict(int)
	pool: dict[str, list[dict]] = defaultdict(list)
	for d in docs:
	ok, why = usable(d)
	if not ok:
	skipped[why] += 1
	continue
	pool[d["category"]].append(d)
	print(f"pool: {sum(len(v) for v in pool.values()):,} usable documents")
	for why, n in sorted(skipped.items(), key=lambda kv: -kv[1]):
	print(f" skipped {n:5} {why[:96]}")

	# ------------------------------------------------------------ calib_train
	spec = recipe["calib_train"]
	target = int(spec["target_tokens"])
	shares = spec["shares"]
	total_share = sum(shares.values())
	# No single document may dominate its slice. An amalgamated single-header or
	# a generated search index is one file, and left uncapped one of them takes
	# 30-60% of a category -- the imatrix then describes that file rather than
	# the category. Overridable per category because a longctx document is
	# supposed to be large.
	frac = dict(spec.get("max_doc_fraction") or {})
	default_frac = float(frac.pop("default", 0.05))
	chosen: list[tuple[dict, str, int]] = []
	actual: dict[str, int] = {}
	oversize: list[tuple[str, str, str, int]] = []
	print(f"\ncalib_train: target {target:,} tokens")
	for cat in sorted(shares):
	budget = int(target * shares[cat] / total_share)
	cap = int(budget * float(frac.get(cat, default_frac)))
	cands = seeded_order(pool.get(cat, []), seed + sum(map(ord, cat)))
	got, taken, skipped_big = 0, 0, 0
	for d in cands:
	if got >= budget:
	break
	text = render_chat(d) if d.get("render") == "chat" else d["text"]
	n = tok.count(text)
	if n == 0:
	continue
	if n > cap:
	skipped_big += 1
	oversize.append((cat, d["source"], d["path"], n))
	continue
	chosen.append((d, text, n))
	got += n
	taken += 1
	actual[cat] = got
	avail = "" if got >= budget * 0.98 else " <-- pool exhausted"
	big = f" ({skipped_big} over the {cap:,}-token cap)" if skipped_big else ""
	print(f" {cat:13} target {budget:9,} got {got:9,} in {taken:5} docs{avail}{big}")
	if oversize:
	print(f"\n {len(oversize)} documents skipped as too large for their slice:")
	for cat, src, path, n in sorted(oversize, key=lambda t: -t[3])[:8]:
	print(f" {cat:12} {n:8,} tok {src[:16]:16} {path[:52]}")

	chosen = sorted(chosen, key=lambda t: (t[0]["category"], t[0]["id"]))
	random.Random(seed).shuffle(chosen)
	total_tokens = sum(n for _, _, n in chosen)
	body = DOC_SEP.join(t for _, t, _ in chosen) + "\n"
	train_path = os.path.join(out_dir, "calib_train.txt")
	with open(train_path, "w", encoding="utf-8") as f:
	f.write(body)
	print(f" wrote {train_path} {len(chosen):,} docs {total_tokens:,} tokens "
	f"({len(body):,} chars)")

	with open(os.path.join(out_dir, "calib_train.manifest.jsonl"), "w",
	encoding="utf-8") as f:
	for d, text, n in chosen:
	f.write(json.dumps({
	"id": d["id"], "category": d["category"], "source": d["source"],
	"license": d["license"], "path": d["path"], "lang": d["lang"],
	"origin": d["origin"], "synthetic": d.get("synthetic", False),
	"render": d.get("render", "text"),
	"reasoning_strength": d.get("reasoning_strength", ""),
	"tokens": n, "chars": len(text), "shard": d["_shard"],
	}, ensure_ascii=False) + "\n")

	# ---------------------------------------------------------- calib_longctx
	lspec = recipe.get("calib_longctx")
	long_docs: list[tuple[dict, str, int]] = []
	if lspec:
	lo, hi = int(lspec["doc_tokens_min"]), int(lspec["doc_tokens_max"])
	ltarget = int(lspec["target_tokens"])
	used = {d["id"] for d, _, _ in chosen}
	cands = seeded_order(
	[d for c in lspec.get("sources", ["longctx"]) for d in pool.get(c, [])
	if d["id"] not in used], seed + 977)
	got = 0
	print(f"\ncalib_longctx: target {ltarget:,} tokens, "
	f"documents of {lo:,}-{hi:,} tokens, unbroken")
	for d in cands:
	if got >= ltarget:
	break
	text = render_chat(d) if d.get("render") == "chat" else d["text"]
	cut = truncate_to_tokens(text, tok, lo, hi)
	if cut is None:
	continue
	n = tok.count(cut)
	long_docs.append((d, cut, n))
	got += n
	lpath = os.path.join(out_dir, "calib_longctx.txt")
	with open(lpath, "w", encoding="utf-8") as f:
	f.write(DOC_SEP.join(t for _, t, _ in long_docs) + "\n")
	lens = sorted(n for _, _, n in long_docs)
	print(f" wrote {lpath} {len(long_docs)} docs {got:,} tokens "
	f"min {lens[0]:,} max {lens[-1]:,}" if lens else " no documents qualified")
	with open(os.path.join(out_dir, "calib_longctx.manifest.jsonl"), "w",
	encoding="utf-8") as f:
	for d, text, n in long_docs:
	f.write(json.dumps({
	"id": d["id"], "category": d["category"], "source": d["source"],
	"license": d["license"], "path": d["path"], "origin": d["origin"],
	"tokens": n, "chars": len(text), "shard": d["_shard"],
	}, ensure_ascii=False) + "\n")

	# ----------------------------------------------------------------- manifest
	shard_hashes = {}
	for p in P.shard_paths(P.POOL_ROOT):
	shard_hashes[os.path.relpath(p, args.repo)] = {
	"sha256": P.sha256_file(p),
	"bytes": os.path.getsize(p),
	"documents": sum(1 for _ in P.read_jsonl(p)),
	}

	lens = sorted(n for _, _, n in chosen)

	def pct(q: float) -> int:
	return lens[min(len(lens) - 1, int(q * len(lens)))] if lens else 0

	by_cat_docs = defaultdict(int)
	by_cat_tok = defaultdict(int)
	synth_tok = 0
	for d, _, n in chosen:
	by_cat_docs[d["category"]] += 1
	by_cat_tok[d["category"]] += n
	if d.get("synthetic"):
	synth_tok += n

	manifest = {
	"build": name,
	"built_at": recipe.get("chat", {}).get("current_date"),
	"build_tool_version": BUILD_VERSION,
	"recipe": {"path": os.path.relpath(args.recipe, args.repo),
	"sha256": P.sha256_file(args.recipe), "seed": seed},
	"model": model,
	"tokenizer": {
	"path": model["tokenizer"],
	"sha256": P.sha256_file(model["tokenizer"]),
	"rows": tok.n_vocab,
	"pre_tokenizer": model.get("pre_tokenizer", "llama4"),
	},
	"llama_cpp": {
	"commit": model.get("llama_cpp_commit", ""),
	"note": "llama-imatrix must be given --parse-special or the chat "
	"markup in calib_train.txt is tokenised as literal text "
	"(parse_special defaults to false in that tool)",
	},
	"calib_train": {
	"path": "calib_train.txt",
	"sha256": P.sha256_file(train_path),
	"bytes": os.path.getsize(train_path),
	"documents": len(chosen),
	"tokens": total_tokens,
	"target_tokens": target,
	"document_separator": DOC_SEP,
	"shares_requested_percent": {k: round(100.0 * v / total_share, 3)
	for k, v in sorted(shares.items())},
	"shares_actual_percent": {k: round(100.0 * by_cat_tok[k] / total_tokens, 3)
	for k in sorted(by_cat_tok)},
	"documents_by_category": dict(sorted(by_cat_docs.items())),
	"tokens_by_category": dict(sorted(by_cat_tok.items())),
	"synthetic_tokens": synth_tok,
	"synthetic_percent": round(100.0 * synth_tok / total_tokens, 3),
	"document_tokens": {
	"p50": pct(0.50), "p90": pct(0.90), "p95": pct(0.95),
	"p99": pct(0.99), "max": lens[-1] if lens else 0,
	"min": lens[0] if lens else 0,
	},
	},
	"pool_sources": dict(sorted(shard_hashes.items())),
	}
	if long_docs:
	llens = sorted(n for _, _, n in long_docs)
	manifest["calib_longctx"] = {
	"path": "calib_longctx.txt",
	"sha256": P.sha256_file(os.path.join(out_dir, "calib_longctx.txt")),
	"documents": len(long_docs),
	"tokens": sum(llens),
	"document_tokens": {"min": llens[0], "p50": llens[len(llens) // 2],
	"max": llens[-1]},
	"contiguous": True,
	}
	with open(os.path.join(out_dir, "manifest.json"), "w", encoding="utf-8") as f:
	json.dump(manifest, f, indent=2, ensure_ascii=False)
	f.write("\n")
	print(f"\nwrote {os.path.join(out_dir, 'manifest.json')}")
	print(json.dumps(manifest["calib_train"]["shares_actual_percent"], indent=2))
	return 0


	if __name__ == "__main__":
	sys.exit(main())