Datasets:

josh-hud
/

code-pii

#!/usr/bin/env python3
"""Synthetic secret-injection dataset generator with cross-document coreference (v3).
Streams permissively-licensed files from codeparrot/github-code-clean (parquet shards
on the Hugging Face Hub), groups files into clusters (real multi-file repos, synthetic
"bundles" of unrelated files, or singletons), gives every cluster an entity pool of
secrets / emails / synthetic persons, injects format-correct synthetic secrets at
positions drawn from Samsung/CredData placement priors (priors.json) with entities
recurring across the files of a cluster, injects labeled hard negatives, and writes
span-labeled JSONL plus a per-cluster entity table (gold replacements included).
Examples
  # 2 GB using the first shards, 4 worker processes, downloading shards on the fly
  python generate_dataset.py --out /data/poc --target-bytes 2e9 --workers 4
  # scale: ~40 GB
  python generate_dataset.py --out /data/full --target-bytes 40e9 --workers 8 --start-shard 0
  # use pre-downloaded shards
  python generate_dataset.py --out /data/poc --shard-dir ./shards --target-bytes 2e9
Dependencies: python>=3.9, pyarrow, faker (stdlib otherwise).
"""
import argparse, base64, hashlib, json, os, random, re, string, sys, time, unicodedata, uuid, urllib.request
from multiprocessing import Pool

try:
    import pyarrow.parquet as pq
    import pyarrow.compute as pc
except ImportError:
    sys.exit("pip install pyarrow")
try:
    from faker import Faker
except ImportError:
    sys.exit("pip install faker")

HF_URL = "https://huggingface.co/datasets/codeparrot/github-code-clean/resolve/main/data/train-{i:05d}-of-00880.parquet"
N_SHARDS = 880
PERMISSIVE = {"mit", "apache-2.0", "bsd-3-clause", "bsd-2-clause", "isc", "unlicense", "cc0-1.0"}

# --------------------------------------------------------------------------------------
# Language syntax table: comment prefix, assignment template, quote char, extra forms
# --------------------------------------------------------------------------------------
# assign: how a hardcoded variable looks. {n}=name {q}=quote {v}=value
LANG = {
    "Python":      dict(c="#", assign=["{n} = {q}{v}{q}", "{N} = {q}{v}{q}", "self.{n} = {q}{v}{q}", "os.environ[{q}{N}{q}] = {q}{v}{q}"], q=["'", '"'], kw=["{q}{n}{q}: {q}{v}{q},", "{n}={q}{v}{q},"]),
    "JavaScript":  dict(c="//", assign=["const {n} = {q}{v}{q};", "let {n} = {q}{v}{q};", "var {n} = {q}{v}{q};", "const {N} = {q}{v}{q};", "process.env.{N} = {q}{v}{q};"], q=["'", '"', "`"], kw=["{n}: {q}{v}{q},", "{q}{n}{q}: {q}{v}{q},"]),
    "TypeScript":  dict(c="//", assign=["const {n} = {q}{v}{q};", "private {n} = {q}{v}{q};", "const {N}: string = {q}{v}{q};", "export const {N} = {q}{v}{q};"], q=["'", '"'], kw=["{n}: {q}{v}{q},"]),
    "Java":        dict(c="//", assign=["String {n} = {q}{v}{q};", "private static final String {N} = {q}{v}{q};", "final String {n} = {q}{v}{q};"], q=['"'], kw=["props.setProperty({q}{n}{q}, {q}{v}{q});"]),
    "C#":          dict(c="//", assign=["string {n} = {q}{v}{q};", "private const string {N} = {q}{v}{q};", "var {n} = {q}{v}{q};"], q=['"'], kw=["[{q}{n}{q}] = {q}{v}{q},"]),
    "GO":          dict(c="//", assign=["{n} := {q}{v}{q}", "const {n} = {q}{v}{q}", "var {n} = {q}{v}{q}", "{N} = {q}{v}{q}"], q=['"', "`"], kw=["{q}{n}{q}: {q}{v}{q},", "{n}: {q}{v}{q},"]),
    "PHP":         dict(c="//", assign=["${n} = {q}{v}{q};", "define({q}{N}{q}, {q}{v}{q});", "$config[{q}{n}{q}] = {q}{v}{q};"], q=["'", '"'], kw=["{q}{n}{q} => {q}{v}{q},"]),
    "Ruby":        dict(c="#", assign=["{n} = {q}{v}{q}", "{N} = {q}{v}{q}", "@{n} = {q}{v}{q}", "ENV[{q}{N}{q}] = {q}{v}{q}"], q=["'", '"'], kw=["{n}: {q}{v}{q},", ":{n} => {q}{v}{q},"]),
    "C":           dict(c="//", assign=["const char *{n} = {q}{v}{q};", "static const char {n}[] = {q}{v}{q};", "#define {N} {q}{v}{q}"], q=['"'], kw=[]),
    "C++":         dict(c="//", assign=["const std::string {n} = {q}{v}{q};", "std::string {n} = {q}{v}{q};", "#define {N} {q}{v}{q}", "static const char* {n} = {q}{v}{q};"], q=['"'], kw=[]),
    "Rust":        dict(c="//", assign=["let {n} = {q}{v}{q};", "const {N}: &str = {q}{v}{q};", "static {N}: &str = {q}{v}{q};"], q=['"'], kw=[]),
    "Scala":       dict(c="//", assign=["val {n} = {q}{v}{q}", "private val {n} = {q}{v}{q}"], q=['"'], kw=["{q}{n}{q} -> {q}{v}{q},"]),
    "Kotlin":      dict(c="//", assign=["val {n} = {q}{v}{q}", "const val {N} = {q}{v}{q}"], q=['"'], kw=[]),
    "Shell":       dict(c="#", assign=["{N}={q}{v}{q}", "export {N}={q}{v}{q}", "{n}={q}{v}{q}"], q=['"', "'", ""], kw=["--{n} {q}{v}{q} \\"]),
    "PowerShell":  dict(c="#", assign=["${n} = {q}{v}{q}", "$env:{N} = {q}{v}{q}"], q=['"', "'"], kw=[]),
    "Batchfile":   dict(c="REM", assign=["set {N}={v}", "SET {N}={v}"], q=[""], kw=[]),
    "Perl":        dict(c="#", assign=["my ${n} = {q}{v}{q};", "${n} = {q}{v}{q};"], q=["'", '"'], kw=["{n} => {q}{v}{q},"]),
    "Lua":         dict(c="--", assign=["local {n} = {q}{v}{q}", "{n} = {q}{v}{q}"], q=['"', "'"], kw=["{n} = {q}{v}{q},"]),
    "Julia":       dict(c="#", assign=["{n} = {q}{v}{q}", "const {N} = {q}{v}{q}"], q=['"'], kw=[]),
    "Haskell":     dict(c="--", assign=["{n} = {q}{v}{q}"], q=['"'], kw=[]),
    "SQL":         dict(c="--", assign=["SET @{n} = {q}{v}{q};", "IDENTIFIED BY {q}{v}{q};", "CREATE USER '{u}'@'%' IDENTIFIED BY {q}{v}{q};"], q=["'"], kw=[]),
    "Dockerfile":  dict(c="#", assign=["ENV {N}={v}", "ENV {N} {v}", "ARG {N}={v}"], q=[""], kw=[]),
    "Makefile":    dict(c="#", assign=["{N} = {v}", "{N} := {v}", "export {N}={v}"], q=[""], kw=[]),
    "CMake":       dict(c="#", assign=["set({N} {q}{v}{q})"], q=['"'], kw=[]),
    "HTML":        dict(c="<!--", assign=['<input type="hidden" name="{n}" value="{v}">', '<meta name="{n}" content="{v}">', "var {n} = {q}{v}{q};"], q=['"'], kw=[]),
    "CSS":         dict(c="/*", assign=["/* {n}: {v} */"], q=[""], kw=[]),
    "Markdown":    dict(c="", assign=["`{n}={v}`", "{n}: `{v}`", "    {N}={v}", "- **{n}**: `{v}`"], q=[""], kw=[]),
    "TeX":         dict(c="%", assign=["\\newcommand{{\\{n}}}{{{v}}}"], q=[""], kw=[]),
    "Visual Basic": dict(c="'", assign=["Dim {n} As String = {q}{v}{q}", "Const {N} As String = {q}{v}{q}"], q=['"'], kw=[]),
    "FORTRAN":     dict(c="!", assign=["character(len=*), parameter :: {n} = {q}{v}{q}"], q=["'", '"'], kw=[]),
    "Assembly":    dict(c=";", assign=["{n} db {q}{v}{q}, 0"], q=['"'], kw=[]),
}
GENERIC = dict(c="#", assign=["{n} = {q}{v}{q}", "{N}={v}"], q=['"', ""], kw=[])

# --------------------------------------------------------------------------------------
# Value generators.  Each returns (value, value_type).
# --------------------------------------------------------------------------------------
R = random.Random()
fake = Faker()
HEX = "0123456789abcdef"
B62 = string.ascii_letters + string.digits
B64 = B62 + "+/"
B64URL = B62 + "-_"

def rs(alpha, n):
    return "".join(R.choice(alpha) for _ in range(n))

def b64(n):
    return base64.b64encode(os.urandom(n)).decode()

def b64url(n):
    raw = n if isinstance(n, bytes) else os.urandom(n)
    return base64.urlsafe_b64encode(raw).decode().rstrip("=")

JWT_NAME = None   # set by the entity pool so an owned JWT carries its owner's name in the payload (no span: the payload is opaque)

def gen_jwt():
    hdr = b64url(json.dumps({"alg": R.choice(["HS256", "RS256"]), "typ": "JWT"}).encode())
    pl = b64url(json.dumps({"sub": str(R.randint(1000, 999999)), "name": JWT_NAME or fake.name(), "iat": R.randint(1500000000, 1750000000), "exp": R.randint(1750000000, 1900000000)}).encode())
    return f"{hdr}.{pl}.{b64url(32)}"

def gen_password():
    """Password sub-generators spanning CredData's charset buckets (lower 49%, digit+lower 16%, mixed 10%, ...).
    The final mix is set by charset-targeted sampling (see gen_value), not by the branch probabilities here."""
    k = R.random()
    if k < 0.30:   # pure lowercase: dictionary words, as in most real leaked passwords
        return R.choice([fake.word() + fake.word(), fake.word() + R.choice(["pass", "pwd", "admin", "secret", "test", "root"]), fake.user_name().replace(".", "").replace("_", ""), rs(string.ascii_lowercase, R.randint(5, 12))])
    if k < 0.45:   # lowercase + digits
        return R.choice([fake.word() + str(R.randint(1, 9999)), rs(string.ascii_lowercase + string.digits, R.randint(6, 12)), fake.word() + str(R.randint(10, 99)) + fake.word()])
    if k < 0.55:   # lowercase + symbol
        return R.choice([fake.word() + R.choice(["_", "-", ".", "!", "@"]) + fake.word(), fake.word() + R.choice(["!", "#", "$"])])
    if k < 0.70:   # mixed case, digits, symbols (Faker-style strong passwords)
        return fake.password(length=R.randint(8, 16), special_chars=R.random() < 0.6)
    if k < 0.85:   # capitalised word + digits (+ symbol)
        return R.choice(["", "", "admin", "root", "dev", "test", "db", "qa"]) + fake.word().capitalize() + R.choice(["", "", "!", "#", "@"]) + str(R.randint(1, 2024))
    if k < 0.92:
        return fake.word() + "_" + fake.word() + str(R.randint(10, 99))
    if k < 0.96:
        return str(R.randint(10 ** 5, 10 ** 9))                       # digits only
    return rs(string.ascii_uppercase, R.randint(6, 10)) if R.random() < 0.5 else fake.word().capitalize() + fake.word().capitalize()

def gen_pem():
    kind = R.choice(["RSA PRIVATE KEY", "RSA PRIVATE KEY", "PRIVATE KEY", "EC PRIVATE KEY", "OPENSSH PRIVATE KEY", "DSA PRIVATE KEY"])
    nbytes = {"EC PRIVATE KEY": R.randint(100, 160), "OPENSSH PRIVATE KEY": R.randint(300, 420)}.get(kind, R.randint(600, 1200))
    body = b64(nbytes)
    lines = [body[i:i + 64] for i in range(0, len(body), 64)]
    return f"-----BEGIN {kind}-----\n" + "\n".join(lines) + f"\n-----END {kind}-----"

def gen_email():
    k = R.random()
    if k < 0.5:
        return fake.free_email()
    if k < 0.75:
        return fake.user_name() + "@" + fake.domain_name()
    if k < 0.9:
        return fake.company_email()
    return fake.first_name().lower() + "." + fake.last_name().lower() + "@" + R.choice(["gmail.com", "outlook.com", "yahoo.com", "protonmail.com", "qq.com"])

# label -> list of (weight, value_type, fn)
GENERATORS = {
    "api_key": [
        (3, "aws_access_key_id", lambda: "AKIA" + rs(string.ascii_uppercase + string.digits, 16)),
        (3, "aws_secret_access_key", lambda: rs(B64, 40)),
        (2, "google_api_key", lambda: "AIza" + rs(B64URL, 35)),
        (2, "stripe_secret_key", lambda: R.choice(["sk_live_", "sk_test_", "rk_live_"]) + rs(B62, 24)),
        (1, "sendgrid_api_key", lambda: "SG." + rs(B64URL, 22) + "." + rs(B64URL, 43)),
        (1, "twilio_api_key", lambda: "SK" + rs(HEX, 32)),
        (1, "mailgun_api_key", lambda: "key-" + rs(HEX, 32)),
        (1, "openai_api_key", lambda: "sk-" + R.choice(["", "proj-"]) + rs(B62, 48)),
        (1, "anthropic_api_key", lambda: "sk-ant-api03-" + rs(B64URL, 93) + "AA"),
        (1, "azure_client_secret", lambda: rs(B62 + "~._-", 40)),
        (6, "generic_hex_key", lambda: rs(HEX, R.choice([32, 32, 40, 48, 64, 64]))),
        (2, "generic_hex_key_upper", lambda: rs(HEX.upper(), R.choice([32, 40, 64]))),
        (4, "generic_base64_secret", lambda: b64(R.choice([24, 32, 32, 48, 64]))),
        (2, "generic_alnum_key", lambda: rs(B62, R.choice([20, 24, 32, 40]))),
        (1, "generic_lower_key", lambda: rs(string.ascii_lowercase, R.choice([16, 20, 24, 32]))),
        (1, "generic_upper_key", lambda: rs(string.ascii_uppercase + string.digits, R.choice([16, 20, 25, 32]))),
        (1, "generic_digit_key", lambda: rs(string.digits, R.choice([12, 16, 20, 32]))),
        (1, "generic_hex_dashed_key", lambda: "-".join(rs(HEX, 8) for _ in range(R.choice([3, 4, 5])))),
    ],
    "token": [
        (3, "github_pat", lambda: R.choice(["ghp_", "gho_", "ghs_"]) + rs(B62, 36)),
        (1, "github_fine_grained_pat", lambda: "github_pat_" + rs(B62, 22) + "_" + rs(B62, 59)),
        (2, "slack_token", lambda: R.choice(["xoxb-", "xoxp-", "xoxa-2-"]) + str(R.randint(10 ** 10, 10 ** 12)) + "-" + str(R.randint(10 ** 10, 10 ** 12)) + "-" + rs(B62, 24)),
        (1, "slack_webhook", lambda: "https://hooks.slack.com/services/T" + rs(string.ascii_uppercase + string.digits, 8) + "/B" + rs(string.ascii_uppercase + string.digits, 8) + "/" + rs(B62, 24)),
        (4, "jwt", gen_jwt),
        (1, "gitlab_pat", lambda: "glpat-" + rs(B62 + "_-", 20)),
        (1, "npm_token", lambda: "npm_" + rs(B62, 36)),
        (1, "discord_bot_token", lambda: rs(B64URL, 24) + "." + rs(B64URL, 6) + "." + rs(B64URL, 27)),
        (1, "digitalocean_pat", lambda: "dop_v1_" + rs(HEX, 64)),
        (1, "hf_token", lambda: "hf_" + rs(B62, 34)),
        (3, "bearer_token_opaque", lambda: rs(B64URL, R.choice([32, 40, 43, 64]))),
        (2, "basic_auth_b64", lambda: base64.b64encode(f"{fake.user_name()}:{gen_password()}".encode()).decode()),
        (3, "generic_hex_token", lambda: rs(HEX, R.choice([32, 40, 48, 64]))),
        (2, "generic_alnum_token", lambda: rs(B62, R.choice([24, 32, 40, 48]))),
        (2, "generic_dashed_token", lambda: "-".join(rs(string.ascii_lowercase + string.digits, R.choice([4, 6, 8])) for _ in range(R.choice([3, 4, 5, 6])))),
        (1, "generic_upper_token", lambda: rs(string.ascii_uppercase + string.digits, R.choice([20, 24, 32]))),
    ],
    "password": [(1, "password", gen_password)],
    "private_key": [(1, "pem_private_key", gen_pem)],
    "email": [(1, "email", gen_email)],
}

def pick_weighted(items):
    tot = sum(w for w, *_ in items)
    r = R.uniform(0, tot)
    for it in items:
        r -= it[0]
        if r <= 0:
            return it
    return items[-1]

def charset(v):
    s = set()
    if re.search(r"[a-z]", v): s.add("lower")
    if re.search(r"[A-Z]", v): s.add("upper")
    if re.search(r"[0-9]", v): s.add("digit")
    if re.search(r"[^A-Za-z0-9]", v): s.add("symbol")
    return "+".join(sorted(s))

def _profile_generators(n=300):
    """Empirical P(charset | generator), used to steer sampling toward CredData's charset mix."""
    prof = {}
    for label, gens in GENERATORS.items():
        for w, vt, fn in gens:
            c = {}
            for _ in range(n):
                cs = charset(fn()); c[cs] = c.get(cs, 0) + 1
            prof[vt] = {k: v / n for k, v in c.items()}
    return prof

GEN_PROFILE = None
CHARSET_TARGET = {}   # label -> {charset: weight}; filled from priors.json by Priors()

def gen_value(label):
    """Pick a generator and value.  When a charset target exists for the label (api_key, password, token),
    first draw a target charset from CredData's distribution, then draw a generator proportionally to
    base_weight * P(charset | generator) and rejection-sample (<=12 tries) until the value matches."""
    global GEN_PROFILE
    gens = GENERATORS[label]
    tgt = CHARSET_TARGET.get(label)
    if not tgt:
        w, vt, fn = pick_weighted(gens)
        return fn(), vt
    if GEN_PROFILE is None:
        GEN_PROFILE = _profile_generators()
    cs = R.choices(list(tgt), list(tgt.values()))[0]
    cands = [(w * GEN_PROFILE[vt].get(cs, 0), vt, fn) for w, vt, fn in gens if GEN_PROFILE[vt].get(cs, 0) > 0]
    if not cands:
        w, vt, fn = pick_weighted(gens)
        return fn(), vt
    w, vt, fn = pick_weighted(cands)
    v = fn()
    for _ in range(12):
        if charset(v) == cs:
            break
        v = fn()
    return v, vt

GEN_BY_VT = {vt: fn for gens in GENERATORS.values() for _, vt, fn in gens}
FIXED_LEN_VT = {"aws_access_key_id", "aws_secret_access_key", "google_api_key", "stripe_secret_key", "sendgrid_api_key", "twilio_api_key",
                "mailgun_api_key", "anthropic_api_key", "azure_client_secret", "github_pat", "github_fine_grained_pat", "gitlab_pat",
                "npm_token", "digitalocean_pat", "hf_token", "discord_bot_token"}
BAD_VALUE_CHARS = set("'\"`")   # the renderer rewrites quotes inside values; entity values must survive rendering unchanged

# --------------------------------------------------------------------------------------
# Persons and entity pool (v3: cross-document coreference)
# --------------------------------------------------------------------------------------
PERSON_LOCALES = ["en_US"] * 7 + ["de_DE", "es_ES", "fr_FR", "it_IT", "pt_BR", "nl_NL", "pl_PL", "sv_SE", "en_GB", "en_IN"]
_FAKERS = {}

def faker_for(locale):
    if locale not in _FAKERS:
        _FAKERS[locale] = Faker(locale)
    return _FAKERS[locale]

def ascii_name(s):
    s = unicodedata.normalize("NFKD", s).encode("ascii", "ignore").decode()
    s = re.sub(r"[^A-Za-z\- ]", "", s).strip(" -")
    return s if re.fullmatch(r"[A-Za-z][A-Za-z\-]*", s) else None

USERNAME_TEMPLATES = {
    "f_last": lambda f, l: (f[0] + l).lower(),
    "first.last": lambda f, l: f"{f}.{l}".lower(),
    "f_last_dd": lambda f, l: (f[0] + l).lower() + "{dd}",
    "first_l": lambda f, l: (f + l[0]).lower(),
    "last_f": lambda f, l: (l + f[0]).lower(),
    "first_last": lambda f, l: f"{f}_{l}".lower(),
}
FREE_DOMAINS = ["gmail.com", "outlook.com", "yahoo.com", "protonmail.com", "hotmail.com", "icloud.com", "gmx.de", "qq.com"]
PERSON_FORMS = ["full_name", "first_name", "last_name", "username", "handle", "email"]
PERSON_FORM_LABEL = {"full_name": "person_name", "first_name": "person_name", "last_name": "person_name", "username": "username", "handle": "username", "email": "email"}
PERSON_FORM_W = {"full_name": 40, "username": 25, "email": 20, "first_name": 10, "last_name": 3, "handle": 2}
PERSON_NAMES = {   # variable names per surface form (assignment / json / yaml contexts)
    "full_name": {"author": 30, "maintainer": 10, "owner": 10, "name": 10, "created_by": 8, "contact": 5, "reviewer": 5, "full_name": 5, "developer": 4, "display_name": 4},
    "first_name": {"first_name": 20, "name": 10, "owner": 5, "assignee": 5, "contact": 3},
    "last_name": {"last_name": 20, "surname": 5, "family_name": 5},
    "username": {"user": 20, "username": 20, "owner": 10, "maintainer": 10, "created_by": 8, "login": 8, "author": 5, "assignee": 5, "user_name": 5, "db_user": 5, "ssh_user": 3},
    "handle": {"handle": 10, "twitter": 10, "github": 10, "owner": 5, "author": 5, "slack": 3},
    "email": {"email": 30, "author_email": 10, "contact": 10, "maintainer_email": 8, "owner_email": 5, "user_email": 5, "from_email": 5, "reply_to": 3},
}
PERSON_COMMENT_TMPL = {
    "full_name": ["{c} Author: {v}", "{c} @author {v}", "{c} Copyright (c) {y} {v}", "{c} Maintainer: {v}", "{c} Written by {v}", "{c} Reviewed by {v}", "{c} Contact: {v}", "{c} created by {v} on {d}"],
    "first_name": ["{c} TODO({v}): clean this up", "{c} ask {v} before changing this", "{c} thanks {v}", "{c} FIXME {v}: handle the error case", "{c} {v}: this is a workaround"],
    "last_name": ["{c} see {v} et al.", "{c} per {v}'s review", "{c} Mr. {v} owns this module"],
    "username": ["{c} TODO({v}): clean this up", "{c} owner: {v}", "{c} reviewed-by: {v}", "{c} login as {v}", "{c} maintainer {v}", "{c} assigned to {v}", "{c} run as user {v}"],
    "handle": ["{c} ping {v} if this breaks", "{c} cc {v}", "{c} owner {v}", "{c} reviewer: {v}"],
    "email": ["{c} Contact: {v}", "{c} Author email: {v}", "{c} report bugs to {v}", "{c} maintainer: {v}", "{c} {v}"],
}

def make_person(repl_of=None):
    """Synthetic person identity: dict of surface forms + the templates that derive them (so a replacement persona
    can be derived with the same templates). repl_of: existing person whose templates/locale/domain are reused."""
    for _ in range(50):
        loc = repl_of["locale"] if repl_of else R.choice(PERSON_LOCALES)
        fk = faker_for(loc)
        first, last = ascii_name(fk.first_name()), ascii_name(fk.last_name())
        if first and last and first.lower() != last.lower() and 2 <= len(first) <= 16 and 2 <= len(last) <= 20:
            break
    else:
        first, last, loc = "Alex", "Morgan", "en_US"
    t = repl_of["templates"] if repl_of else {
        "username": R.choice(list(USERNAME_TEMPLATES)), "dd": f"{R.randint(0, 99):02d}",
        "email": R.choice(["username@domain", "username@domain", "first.last@free", "first.last@domain"]),
        "domain": R.choice([fake.domain_name(), fake.domain_name(), fake.free_email_domain()]), "free": R.choice(FREE_DOMAINS)}
    uname = USERNAME_TEMPLATES[t["username"]](first, last).replace("{dd}", t["dd"]).replace("-", "")
    fl = f"{first}.{last}".lower().replace("-", "")
    email = {"username@domain": f"{uname}@{t['domain']}", "first.last@free": f"{fl}@{t['free']}", "first.last@domain": f"{fl}@{t['domain']}"}[t["email"]]
    forms = {"full_name": f"{first} {last}", "first_name": first, "last_name": last, "username": uname, "handle": "@" + uname, "email": email}
    return {"forms": forms, "templates": t, "locale": loc}

def same_shape(value):
    """Random string with the same length and per-position character class. Hex-only values stay hex (lower or upper);
    otherwise lower/upper/digit positions keep their class and symbols are copied."""
    low = R.choice
    if re.fullmatch(r"[0-9a-f]+", value) and re.search(r"[a-f]", value):
        return "".join(low("abcdef") if ch.isalpha() else low(string.digits) for ch in value)
    if re.fullmatch(r"[0-9A-F]+", value) and re.search(r"[A-F]", value):
        return "".join(low("ABCDEF") if ch.isalpha() else low(string.digits) for ch in value)
    return "".join(low(string.ascii_lowercase) if ch.islower() else low(string.ascii_uppercase) if ch.isupper() else low(string.digits) if ch.isdigit() else ch for ch in value)

VT_PREFIX = {   # vendor formats: the prefix is part of the format and must survive replacement
    "aws_access_key_id": ["AKIA"], "google_api_key": ["AIza"], "stripe_secret_key": ["sk_live_", "sk_test_", "rk_live_"], "sendgrid_api_key": ["SG."],
    "twilio_api_key": ["SK"], "mailgun_api_key": ["key-"], "openai_api_key": ["sk-proj-", "sk-"], "anthropic_api_key": ["sk-ant-api03-"],
    "github_pat": ["ghp_", "gho_", "ghs_"], "github_fine_grained_pat": ["github_pat_"], "slack_token": ["xoxb-", "xoxp-", "xoxa-2-"],
    "slack_webhook": ["https://hooks.slack.com/services/T"], "gitlab_pat": ["glpat-"], "npm_token": ["npm_"], "digitalocean_pat": ["dop_v1_"], "hf_token": ["hf_"],
}
VT_SUFFIX = {"anthropic_api_key": "AA"}

def gen_replacement(vt, value, label):
    """A fresh value of the same value_type and format, different from value: same length and per-position character
    class (vendor prefixes/suffixes kept, hex stays hex), PEMs keep header kind and line count, JWTs are regenerated
    (three base64url segments), passwords and emails are redrawn from their generator with the same charset class."""
    if vt == "pem_private_key":
        parts = value.split("\n"); body = "".join(parts[1:-1]); pad = len(body) - len(body.rstrip("="))
        body = rs(B64, len(body) - pad) + "=" * pad
        return parts[0] + "\n" + "\n".join(body[i:i + 64] for i in range(0, len(body), 64)) + "\n" + parts[-1]
    if vt == "jwt":
        return gen_jwt()
    if vt in ("password", "email"):
        fn = GEN_BY_VT[vt]; cs = charset(value); best = None
        for _ in range(8):
            v = fn()
            if v == value or BAD_VALUE_CHARS & set(v): continue
            if charset(v) == cs and abs(len(v) - len(value)) <= 3: return v
            if best is None and charset(v) == cs: best = v
        return best or same_shape(value)
    pre = next((p for p in VT_PREFIX.get(vt, []) if value.startswith(p)), "")
    suf = VT_SUFFIX.get(vt, "") if value.endswith(VT_SUFFIX.get(vt, "\0")) else ""
    core = value[len(pre):len(value) - len(suf)]
    for _ in range(3):
        v = pre + same_shape(core) + suf
        if v != value: return v
    return v

class EntityPool:
    """Per-cluster registry of entities (secrets, emails, persons) with canonical and replacement surface forms."""
    def __init__(self, cluster_id, n_persons):
        self.cid = cluster_id; self.entities = []; self.by_label = {}; self.persons = []; self.secrets = []
        self.used = set()                       # every original and replacement string in the cluster (uniqueness)
        for _ in range(n_persons):
            self.new_person()
        self.first_name_unique = {}
        for p in self.persons:
            f = p["forms"]["first_name"].lower()
            self.first_name_unique[f] = self.first_name_unique.get(f, 0) + 1

    def _add(self, e):
        e["entity_id"] = f"{self.cid}/e{len(self.entities):03d}"; e["n_mentions"] = 0; e["docs"] = set()
        self.entities.append(e); self.by_label.setdefault(e["entity_type"], []).append(e)
        if e["entity_type"] != "person": self.secrets.append(e)
        return e

    def new_person(self):
        for _ in range(20):
            p = make_person()
            if not any(v in self.used for v in p["forms"].values()):
                break
        own = {v.lower() for v in p["forms"].values()}
        for _ in range(20):
            rp = make_person(repl_of=p)
            if not any(v.lower() in own or v in self.used for v in rp["forms"].values()):
                break
        self.used.update(p["forms"].values()); self.used.update(rp["forms"].values())
        e = self._add({"entity_type": "person", "value_type": "person", "canonical": p["forms"]["full_name"], "forms": p["forms"],
                       "replacement_canonical": rp["forms"]["full_name"], "replacement_forms": rp["forms"], "owner_entity_id": None,
                       "_templates": p["templates"], "_password": None})
        self.persons.append(e)
        return e

    def person_password(self, person):
        """A password entity owned by the person (used for url credentials / basic auth); created once per person."""
        if person["_password"] is None:
            person["_password"] = self.new_secret("password", owner=person)
        return person["_password"]

    def new_secret(self, label, owner=None):
        global JWT_NAME
        JWT_NAME = owner["forms"]["full_name"] if owner else None
        for _ in range(20):
            if label == "token" and owner and R.random() < 0.15:
                pw = self.person_password(owner)
                value, vt = base64.b64encode(f"{owner['forms']['username']}:{pw['canonical']}".encode()).decode(), "basic_auth_b64"
                repl = base64.b64encode(f"{owner['replacement_forms']['username']}:{pw['replacement_canonical']}".encode()).decode()
            else:
                value, vt = gen_value(label)
                repl = gen_replacement(vt, value, label)
            if value not in self.used and repl not in self.used and value != repl and not (BAD_VALUE_CHARS & set(value)):
                break
        JWT_NAME = None
        self.used.add(value); self.used.add(repl)
        return self._add({"entity_type": label, "value_type": vt, "canonical": value, "forms": {"literal": value},
                          "replacement_canonical": repl, "replacement_forms": {"literal": repl},
                          "owner_entity_id": owner["entity_id"] if owner else None})

    def pick(self, label):
        c = self.by_label.get(label)
        return R.choice(c) if c else None

    def person_form(self, person):
        w = dict(PERSON_FORM_W)
        if self.first_name_unique.get(person["forms"]["first_name"].lower(), 0) > 1:
            w.pop("first_name")
        return R.choices(list(w), list(w.values()))[0]

    def note(self, e, doc_index):
        e["n_mentions"] += 1; e["docs"].add(doc_index)

    def table(self):
        """Entity rows for the cluster. Persons without any mention are dropped and owner links to them cleared."""
        alive = {e["entity_id"] for e in self.entities if e["n_mentions"] > 0}
        out = []
        for e in self.entities:
            if e["entity_id"] not in alive:
                continue
            out.append({"entity_id": e["entity_id"], "entity_type": e["entity_type"], "value_type": e["value_type"], "canonical": e["canonical"],
                        "forms": e["forms"], "replacement_canonical": e["replacement_canonical"], "replacement_forms": e["replacement_forms"],
                        "owner_entity_id": e["owner_entity_id"] if e["owner_entity_id"] in alive else None,
                        "n_mentions": e["n_mentions"], "n_docs": len(e["docs"])})
        return out

# Hard negatives: (value_type, fn, label_of_similar_secret)
HARD_NEGATIVES = [
    ("placeholder_name", lambda: R.choice(["John Doe", "Jane Doe", "Jane Smith", "Your Name", "Anonymous", "unknown", "nobody", "admin", "root", "guest", "test user", "Firstname Lastname", "<your name>", "author name", "user", "someone"])),
    ("placeholder", lambda: R.choice(["CHANGEME", "changeme", "<YOUR_API_KEY>", "YOUR_API_KEY_HERE", "xxxxxxxxxxxxxxxxxxxx", "********", "REPLACE_ME", "TODO", "<password>", "${PASSWORD}", "your-secret-here", "EXAMPLE_KEY", "AKIAIOSFODNN7EXAMPLE", "wJalrXUtnFEMI/K7MDENG/bPxRfiCYEXAMPLEKEY", "sk_test_XXXXXXXXXXXXXXXXXXXXXXXX", "ghp_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", "dummy", "secret", "password", "123456", "test"])),
    ("example_email", lambda: fake.user_name() + "@" + R.choice(["example.com", "example.org", "test.com", "localhost", "domain.com", "email.com", "company.com"])),
    ("uuid", lambda: str(uuid.uuid4()) if R.random() < 0.8 else str(uuid.uuid4()).upper()),
    ("sha256_hash", lambda: hashlib.sha256(os.urandom(8)).hexdigest()),
    ("sha1_git_sha", lambda: hashlib.sha1(os.urandom(8)).hexdigest()),
    ("md5_hash", lambda: hashlib.md5(os.urandom(8)).hexdigest()),
    ("env_var_reference", lambda: R.choice(["os.environ['{N}']", "os.getenv('{N}')", "process.env.{N}", "${{{N}}}", "$({N})", "ENV['{N}']", "System.getenv(\"{N}\")", "{{{{ secrets.{N} }}}}", "${{{{ secrets.{N} }}}}", "Environment.GetEnvironmentVariable(\"{N}\")"]).format(N=R.choice(["API_KEY", "SECRET_KEY", "DB_PASSWORD", "GITHUB_TOKEN", "AWS_SECRET_ACCESS_KEY", "TOKEN", "PASSWORD"]))),
    ("public_key_pem", lambda: "-----BEGIN PUBLIC KEY-----\n" + "\n".join([b64(220)[i:i + 64] for i in range(0, 294, 64)]) + "\n-----END PUBLIC KEY-----"),
    ("ssh_public_key", lambda: "ssh-rsa " + b64(R.randint(150, 279)) + " " + fake.user_name() + "@" + fake.hostname()),
    ("version_or_id", lambda: R.choice([fake.ipv4(), "v" + str(R.randint(0, 9)) + "." + str(R.randint(0, 20)) + "." + str(R.randint(0, 50)), "com.example." + fake.word(), str(R.randint(10 ** 9, 10 ** 12)), "/usr/local/share/" + fake.word()])),
    ("base64_non_secret", lambda: base64.b64encode(fake.sentence().encode()).decode()),
]

# --------------------------------------------------------------------------------------
# Placement logic (driven by priors.json)
# --------------------------------------------------------------------------------------
class Priors:
    def __init__(self, path):
        p = json.load(open(path))
        ld = p["label_distribution"]
        # keep the POC label set; CredData's UUID class becomes a hard negative, 'other' dropped
        self.labels = [l for l in ("api_key", "password", "token", "private_key") if l in ld]
        self.label_w = [ld[l] for l in self.labels]
        # emails are not a CredData class; give them a fixed share
        self.labels.append("email"); self.label_w.append(int(0.12 * sum(self.label_w)))
        self.ctx_by_label = {l: p["context_by_label"].get(l, {"assignment": 1}) for l in self.labels}
        self.ctx_by_label["email"] = {"assignment": 40, "comment": 35, "json_key": 10, "yaml_key": 10}
        self.names = {l: p["var_names_by_label_top40"].get(l, {"secret": 1}) for l in self.labels}
        self.names["email"] = {"email": 30, "admin_email": 10, "user_email": 10, "from_email": 8, "contact": 5, "author_email": 5, "smtp_user": 5, "mail": 4, "username": 4}
        self.names["token"] = dict(self.names["token"]); self.names["token"].update({"token": 50, "access_token": 30, "auth_token": 20, "api_token": 15, "github_token": 10, "slack_token": 8, "bearer": 5})
        self.names["api_key"] = dict(self.names["api_key"]); self.names["api_key"].update({"api_key": 60, "apikey": 30, "secret_key": 30, "aws_secret_access_key": 20, "aws_access_key_id": 20, "client_secret": 20, "private_key": 5, "stripe_key": 5})
        self.names["private_key"] = {"private_key": 30, "key": 20, "rsa_key": 10, "ssh_key": 10, "pem": 5, "signing_key": 5, "server_key": 5}
        global GEN_PROFILE
        if GEN_PROFILE is None:
            GEN_PROFILE = _profile_generators()
        for l in ("api_key", "password", "token"):
            d = p.get("charset_by_label", {}).get(l, {})
            reachable = set().union(*(GEN_PROFILE[vt].keys() for _, vt, _ in GENERATORS[l]))
            CHARSET_TARGET[l] = {cs: n for cs, n in d.items() if cs in reachable and n > 0}
        h = p.get("secrets_per_file_hist", {"1": 60, "2": 20, "3": 10, "4": 5})
        # CredData's last bucket is "10+", dominated by a few credential-dump files; damp it
        self.nsec = [min(int(k), 5) for k in h]; self.nsec_w = [h[k] / (4 if int(k) >= 10 else 1) for k in h]
        self.quote_style = p.get("quote_style", {})

    def sample_label(self):
        return R.choices(self.labels, self.label_w)[0]

    def sample_ctx(self, label, allowed):
        d = {k: v for k, v in self.ctx_by_label[label].items() if k in allowed}
        if not d:
            return "assignment"
        return R.choices(list(d), list(d.values()))[0]

    def sample_name(self, label):
        d = self.names[label]
        return R.choices(list(d), list(d.values()))[0]

    def sample_nsec(self):
        return R.choices(self.nsec, self.nsec_w)[0]

def name_variants(n):
    """Return (snake, UPPER, camel) forms of a variable name."""
    n = re.sub(r"[^a-z0-9_]", "_", n.lower()).strip("_") or "secret"
    parts = n.split("_")
    camel = parts[0] + "".join(x.capitalize() for x in parts[1:])
    return n, n.upper(), camel

PEM_ESCAPE = ("\n", "\\n\" +\n    \"")   # how a PEM is written inside Java/C/C#/C++ string concatenation

def render(ctx, lang, label, name, value, is_negative=False, user=None, comment_tmpls=None):
    """Render one injected line (may be multi-line for PEM).
    Returns (line_text, rel_start, rel_end, form, user_pos): form is 'literal' or 'escaped_literal' (PEM inside a
    string concatenation); user_pos is (start, end) of the user string when `user` was given and the construct
    contains it (url credential, 'login as' comment, SQL CREATE USER), else None."""
    L = LANG.get(lang, GENERIC)
    snake, upper, camel = name_variants(name)
    n = R.choice([snake, snake, camel]) if lang in ("JavaScript", "TypeScript", "Java", "C#", "GO", "Scala", "Kotlin") else snake
    q = R.choice(L["q"]) if L["q"] else '"'
    if value and ("'" in value or '"' in value or " " in value or "<" in value) and q == "":
        q = '"'
    if q and q in value:
        q = "'" if q == '"' else '"'
        if q in value:
            value = value.replace(q, "")
    user_given = user is not None
    _u = [user]
    def U():   # lazy: Faker user_name is comparatively slow and most constructs never use it
        if _u[0] is None: _u[0] = fake.user_name()
        return _u[0]
    form = "literal"
    if label == "private_key" or "\n" in value:
        ctx = "pem_block"
    if ctx == "pem_block":
        if lang in ("Python", "JavaScript", "TypeScript", "Ruby", "GO", "PHP", "Shell"):
            opener = {"Python": f'{n} = """', "JavaScript": f"const {n} = `", "TypeScript": f"const {n} = `", "Ruby": f"{n} = <<~KEY\n", "GO": f"{n} := `", "PHP": f"${n} = <<<KEY\n", "Shell": f"cat > ~/.ssh/id_rsa <<'EOF'\n"}[lang]
            closer = {"Python": '"""', "JavaScript": "`;", "TypeScript": "`;", "Ruby": "\nKEY", "GO": "`", "PHP": "\nKEY;", "Shell": "\nEOF"}[lang]
            line = opener + value + closer
        elif lang in ("Java", "C#", "C", "C++"):
            esc = value.replace(*PEM_ESCAPE)
            line = f'String {n} = "{esc}";' if lang == "Java" else f'const char *{n} = "{esc}";' if lang in ("C", "C++") else f'string {n} = "{esc}";'
            value = esc; form = "escaped_literal"
        else:
            line = value
        s = line.index(value)
        return line, s, s + len(value), form, None
    if ctx == "comment":
        c = L["c"] or "#"
        tmpl = R.choice(comment_tmpls or ["{c} {n}: {v}", "{c} {n} = {v}", "{c} TODO: remove hardcoded {n} {v}", "{c} old {n}: {v}", "{c} {n}={v}", "{c} login as {u} / {v}"])
        line = tmpl.format(c=c, n=n, v=value, u=U(), y=R.randint(2008, 2025), d=fake.date() if "{d}" in tmpl else "")
        if c == "<!--": line += " -->"
        if c == "/*": line += " */"
    elif ctx == "url_credential":
        scheme = R.choice(["postgres", "postgresql", "mysql", "mongodb", "mongodb+srv", "redis", "amqp", "https", "http", "ftp", "sftp"])
        host = R.choice([fake.hostname(), "localhost", fake.ipv4(), fake.domain_name()])
        url = f"{scheme}://{U()}:{value}@{host}" + R.choice([":5432/app", "/", ":6379/0", "/db", "/api/v1", ":3306/prod", ""])
        a = R.choice(L["assign"]) if ctx != "pem_block" else "{n} = {q}{v}{q}"
        if lang == "SQL": a = "{n} = {q}{v}{q}"
        line = a.format(n=R.choice(["database_url", "db_url", "conn_str", "connection_string", "url", "uri", "mongo_uri", "redis_url", "broker_url", "dsn"]) if "{n}" in a else "", N=R.choice(["DATABASE_URL", "DB_URI", "MONGO_URL", "REDIS_URL", "CONNECTION_STRING"]), q=q, v=url, u=U())
    elif ctx == "auth_header":
        if label == "token" and R.random() < 0.6:
            hv = f"Bearer {value}"
        else:
            hv = f"Basic {value}" if re.fullmatch(r"[A-Za-z0-9+/=]+", value) else f"Bearer {value}"
        tmpls = AUTH_TMPL.get(lang, ['{q}Authorization{q}: {q}{hv}{q},', "Authorization: {hv}"])
        tmpl = R.choice(tmpls)
        line = tmpl.format(q=q or '"', hv=hv)
    elif ctx in ("json_key", "yaml_key", "env_assignment", "ini_or_props", "function_arg") and L["kw"] and ctx != "env_assignment":
        line = R.choice(L["kw"]).format(n=n, N=upper, q=(q if q in ("'", '"') else '"'), v=value)
    elif ctx == "env_assignment":
        line = R.choice(["{N}={v}", "{N}={q}{v}{q}", "export {N}={q}{v}{q}", "ENV {N}={v}"]).format(N=upper, q=q, v=value) if lang in ("Shell", "Dockerfile", "Makefile", "Markdown") else R.choice(L["assign"]).format(n=n, N=upper, q=q, v=value, u=U())
    else:
        line = R.choice(L["assign"]).format(n=n, N=upper, q=q, v=value, u=U())
    s = line.rindex(value)
    user_pos = None
    if user_given:
        for m in re.finditer(re.escape(_u[0]), line):
            if m.end() <= s and (m.end() == len(line) or not line[m.end()].isalnum()) and (m.start() == 0 or not line[m.start() - 1].isalnum()):
                user_pos = (m.start(), m.end())
    return line, s, s + len(value), "literal", user_pos

AUTH_TMPL = {
    "Python": ['headers[{q}Authorization{q}] = {q}{hv}{q}', 'headers = {{{q}Authorization{q}: {q}{hv}{q}}}', '{q}Authorization{q}: {q}{hv}{q},'],
    "JavaScript": ['{q}Authorization{q}: {q}{hv}{q},', 'headers.Authorization = {q}{hv}{q};', "xhr.setRequestHeader({q}Authorization{q}, {q}{hv}{q});"],
    "TypeScript": ['{q}Authorization{q}: {q}{hv}{q},', 'headers.set({q}Authorization{q}, {q}{hv}{q});'],
    "GO": ['req.Header.Set({q}Authorization{q}, {q}{hv}{q})', 'req.Header.Add({q}Authorization{q}, {q}{hv}{q})'],
    "Java": ['conn.setRequestProperty({q}Authorization{q}, {q}{hv}{q});', '.header({q}Authorization{q}, {q}{hv}{q})', 'headers.set({q}Authorization{q}, {q}{hv}{q});'],
    "C#": ['client.DefaultRequestHeaders.Add({q}Authorization{q}, {q}{hv}{q});', 'request.Headers[{q}Authorization{q}] = {q}{hv}{q};'],
    "PHP": ["'Authorization: {hv}',", "$headers[{q}Authorization{q}] = {q}{hv}{q};"],
    "Ruby": ["request[{q}Authorization{q}] = {q}{hv}{q}", "headers: {{ {q}Authorization{q} => {q}{hv}{q} }}"],
    "Shell": ["-H {q}Authorization: {hv}{q} \\", "curl -H {q}Authorization: {hv}{q} https://api.example.com/v1/me"],
    "Markdown": ["    Authorization: {hv}", "`Authorization: {hv}`"],
    "Rust": ['.header({q}Authorization{q}, {q}{hv}{q})'],
    "C": ['headers = curl_slist_append(headers, {q}Authorization: {hv}{q});'],
    "C++": ['headers = curl_slist_append(headers, {q}Authorization: {hv}{q});'],
    "PowerShell": ['$headers = @{{ Authorization = {q}{hv}{q} }}'],
}
NEG_NAME_LABEL = {"example_email": "email", "public_key_pem": "private_key", "ssh_public_key": "private_key"}
NEG_NAMES = {
    "private_key": {"public_key": 30, "pubkey": 10, "ssh_public_key": 10, "authorized_key": 5, "verify_key": 5, "cert": 5},
    "hash": {"checksum": 20, "sha256": 10, "sha1": 10, "md5": 10, "digest": 10, "commit": 10, "etag": 5, "integrity": 5, "file_hash": 5, "revision": 5, "content_hash": 5, "ref": 3, "git_sha": 3},
    "uuid": {"id": 20, "request_id": 10, "tenant_id": 10, "client_id": 10, "device_id": 8, "correlation_id": 8, "trace_id": 5, "job_id": 5, "uuid": 10, "guid": 5, "instance_id": 5},
    "version_or_id": {"version": 20, "host": 10, "ip": 10, "app_id": 10, "path": 10, "package": 5, "bundle_id": 5, "build": 5, "server": 5},
    "base64_non_secret": {"payload": 20, "data": 10, "body": 10, "encoded": 10, "message": 10, "content": 5, "blob": 5},
    "person": {"author": 20, "owner": 10, "user": 10, "maintainer": 10, "name": 10, "username": 8, "created_by": 5, "contact": 5},
}
NEG_NAME_GROUP = {"sha256_hash": "hash", "sha1_git_sha": "hash", "md5_hash": "hash", "uuid": "uuid", "version_or_id": "version_or_id", "base64_non_secret": "base64_non_secret", "public_key_pem": "private_key", "ssh_public_key": "private_key", "placeholder_name": "person"}

CTX_ALLOWED = {"assignment", "json_key", "yaml_key", "env_assignment", "comment", "url_credential", "auth_header", "function_arg", "pem_block"}

PERSON_CTX = {"comment": 45, "assignment": 35, "json_key": 10, "yaml_key": 10}
NO_ENTITY = {"entity_id": None, "surface_form_type": None, "replacement": None}

def span_meta(ent, label, vt, ctx, form, surface_key):
    repl = ent["replacement_forms"][surface_key]
    if form == "escaped_literal":
        repl = repl.replace(*PEM_ESCAPE)
    return {"label": label, "value_type": vt, "context": ctx, "synthetic": True, "entity_id": ent["entity_id"], "surface_form_type": form if surface_key == "literal" else surface_key, "replacement": repl}

def inject(text, lang, priors, n_secrets, n_negs, pool, doc_index, cfg, n_person_mentions=0):
    """Insert n_secrets secret lines (entities drawn from / added to the cluster pool), n_person_mentions person
    mentions and n_negs hard-negative lines into text. Returns (text, spans, negs)."""
    lines = text.split("\n")
    nl = len(lines)
    inserts = []  # (line_index, rendered_line, [(rel_start, rel_end, meta), ...])
    used_here = []
    ents = {e["entity_id"]: e for e in pool.entities}
    for _ in range(n_secrets):
        label = priors.sample_label()
        if used_here and R.random() < cfg["p_repeat"]:            # second mention of an entity already in this file
            ent = R.choice(used_here); label = ent["entity_type"]
        elif pool.secrets and R.random() < cfg["p_reuse"]:   # recurrence of a cluster entity (any label), preferably one not yet in this file
            ent = R.choice([e for e in pool.secrets if e not in used_here] or pool.secrets); label = ent["entity_type"]
        else:
            owner = R.choice(pool.persons) if pool.persons and label != "email" and R.random() < cfg["p_owner"] else None
            ent = pool.new_secret(label, owner)
        ctx = priors.sample_ctx(label, CTX_ALLOWED)
        name = priors.sample_name(label)
        owner = ents.get(ent["owner_entity_id"])   # owners are persons, all created before the first document
        line, s, e, form, upos = render(ctx, lang, label, name, ent["canonical"], user=owner["forms"]["username"] if owner else None)
        if label == "private_key": ctx = "pem_block"
        parts = [(s, e, span_meta(ent, label, ent["value_type"], ctx, form, "literal"))]
        pool.note(ent, doc_index); used_here.append(ent)
        if upos:
            parts.append((upos[0], upos[1], span_meta(owner, "username", "username", ctx, "literal", "username")))
            pool.note(owner, doc_index)
        inserts.append((R.randint(0, nl), line, parts))
    for _ in range(n_person_mentions):
        person = R.choice(pool.persons)
        fk = pool.person_form(person)
        ctx = R.choices(list(PERSON_CTX), list(PERSON_CTX.values()))[0]
        name = R.choices(list(PERSON_NAMES[fk]), list(PERSON_NAMES[fk].values()))[0]
        line, s, e, form, _ = render(ctx, lang, PERSON_FORM_LABEL[fk], name, person["forms"][fk], comment_tmpls=PERSON_COMMENT_TMPL[fk])
        inserts.append((R.randint(0, nl), line, [(s, e, span_meta(person, PERSON_FORM_LABEL[fk], "email" if fk == "email" else fk, ctx, "literal", fk))]))
        pool.note(person, doc_index)
    for _ in range(n_negs):
        vt, fn = R.choice(HARD_NEGATIVES)
        value = fn()
        if vt in ("env_var_reference",):
            line = R.choice(LANG.get(lang, GENERIC)["assign"]).format(n=R.choice(["api_key", "password", "token", "secret"]), N=R.choice(["API_KEY", "PASSWORD", "TOKEN"]), q="", v=value, u="u")
            s = line.rindex(value); e = s + len(value)
        else:
            label_like = NEG_NAME_LABEL.get(vt) or R.choice(["api_key", "password", "token"])
            ctx = priors.sample_ctx(label_like, CTX_ALLOWED - {"pem_block", "auth_header", "url_credential"})
            if "\n" in value: ctx = "pem_block"
            grp = NEG_NAME_GROUP.get(vt)
            name = R.choices(list(NEG_NAMES[grp]), list(NEG_NAMES[grp].values()))[0] if grp else priors.sample_name(label_like)
            line, s, e, _, _ = render(ctx, lang, "api_key" if label_like == "private_key" else label_like, name, value, is_negative=True)
        inserts.append((R.randint(0, nl), line, [(s, e, dict({"label": "not_secret", "value_type": vt, "context": None, "synthetic": True}, **NO_ENTITY))]))
    inserts.sort(key=lambda x: x[0])
    out = []; spans = []; negs = []; pos = 0; li = 0
    for idx, line, parts in inserts:
        while li < idx:
            out.append(lines[li]); pos += len(lines[li]) + 1; li += 1
        ref = lines[idx] if idx < nl else (lines[-1] if lines else "")
        indent = re.match(r"[ \t]*", ref).group(0) if lang != "Markdown" else ""
        if "\n" in line and lang not in ("Python",):
            indent = ""
        full = indent + line
        for s, e, meta in parts:
            rec = dict(meta, start=pos + len(indent) + s, end=pos + len(indent) + e)
            (negs if meta["label"] == "not_secret" else spans).append(rec)
        out.append(full); pos += len(full) + 1
    while li < nl:
        out.append(lines[li]); pos += len(lines[li]) + 1; li += 1
    spans.sort(key=lambda r: r["start"]); negs.sort(key=lambda r: r["start"])
    return "\n".join(out), spans, negs

# --------------------------------------------------------------------------------------
# Shard processing
# --------------------------------------------------------------------------------------
def fetch_shard(i, shard_dir, keep):
    os.makedirs(shard_dir, exist_ok=True)
    p = os.path.join(shard_dir, f"train-{i:05d}-of-00880.parquet")
    if os.path.exists(p):
        return p, False
    tmp = p + ".part"
    urllib.request.urlretrieve(HF_URL.format(i=i), tmp)
    os.replace(tmp, p)
    return p, not keep

BUNDLE_SIZE = {2: 35, 3: 25, 4: 15, 5: 10, 6: 7, 7: 4, 8: 4}
N_PERSONS_MULTI = ([0, 1, 2, 3], [30, 40, 20, 10])
N_PERSONS_SINGLE = ([0, 1], [70, 30])
N_PERSON_MENTIONS = ([0, 1, 2], [65, 25, 10])

def rec_id(i, ex):
    return f"ghcc-{i:05d}-" + hashlib.sha1(f"{ex['repo_name']}/{ex['path']}".encode()).hexdigest()[:16]

def plan_clusters(pf, i, cfg, stats):
    """Pass 1 (cheap: no text is kept): decide which kept rows form which cluster.
    Returns (row_index -> cluster_key, {cluster_key: {"kind", "members": [row_index...], "key"}})."""
    rows = []   # (kept, repo, lang, rid) per row of the shard
    for rg in range(pf.metadata.num_row_groups):
        tb = pf.read_row_group(rg, columns=["repo_name", "path", "license", "language", "code"])
        lens = pc.utf8_length(tb["code"]).to_pylist()
        for repo, path, lic, lang, n in zip(tb["repo_name"].to_pylist(), tb["path"].to_pylist(), tb["license"].to_pylist(), tb["language"].to_pylist(), lens):
            if cfg["max_rows"] and len(rows) >= cfg["max_rows"]:
                rows.append(None); continue
            if lic not in PERMISSIVE:
                stats["skipped_license"] += 1; rows.append(None); continue
            if not (cfg["min_bytes"] <= n <= cfg["max_bytes"]):
                stats["skipped_size"] += 1; rows.append(None); continue
            rows.append((repo, lang, rec_id(i, {"repo_name": repo, "path": path})))
    by_repo = {}
    for idx, r in enumerate(rows):
        if r: by_repo.setdefault(r[0], []).append(idx)
    clusters = {}; plan = {}
    def add(kind, key, members):
        ck = f"{kind}:{key}"
        clusters[ck] = {"kind": kind, "key": key, "members": members}
        for m in members: plan[m] = ck
    loose = []
    for repo, idxs in by_repo.items():
        if len(idxs) >= 2 and not cfg["no_repo_clusters"]:
            k = -(-len(idxs) // cfg["cluster_cap"])            # split big repos into near-equal chunks <= cap
            per = -(-len(idxs) // k)
            for c in range(k):
                add("repo", f"{repo}#{c}", idxs[c * per:(c + 1) * per])
        else:
            loose.extend(idxs)
    loose.sort()
    queues = {}   # bundle queues: per language (mono-language bundles) and one mixed; bounded buffering in pass 2
    def flush(qk):
        mem = queues.pop(qk)["m"]
        if len(mem) >= 2:
            add("bundle", "+".join(sorted(rows[m][2] for m in mem)), mem)
        else:
            for m in mem: add("singleton", rows[m][2], [m])
    for idx in loose:
        if R.random() < cfg["singleton_frac"]:
            add("singleton", rows[idx][2], [idx]); continue
        qk = rows[idx][1] if R.random() < 0.5 else "*"
        q = queues.setdefault(qk, {"m": [], "n": R.choices(list(BUNDLE_SIZE), list(BUNDLE_SIZE.values()))[0]})
        q["m"].append(idx)
        if len(q["m"]) >= q["n"]:
            flush(qk)
    for qk in list(queues):
        flush(qk)
    return plan, clusters

def emit_cluster(c, i, cfg, priors, fh, stats):
    kind = c["kind"]
    cid = "c-%05d-%s" % (i, hashlib.blake2b(f"{kind}:{c['key']}".encode(), digest_size=8).hexdigest())
    seed = int.from_bytes(hashlib.blake2b(f"{cfg['seed']}:{cid}".encode(), digest_size=8).digest(), "big")
    R.seed(seed); Faker.seed(seed)
    members = [c["buf"][m] for m in c["members"]]
    n = len(members)
    np_ = 0 if cfg["no_persons"] else R.choices(*(N_PERSONS_MULTI if n > 1 else N_PERSONS_SINGLE))[0]
    pool = EntityPool(cid, np_)
    recs = []
    for doc_index, ex in enumerate(members):
        text, lang = ex["code"], ex["language"]
        spans, negs = [], []
        if R.random() < cfg["inject_frac"]:
            n_sec = priors.sample_nsec() if R.random() < 0.85 else 0
            n_neg = R.choice([0, 0, 1, 1, 1, 2, 2, 3])
            n_pm = R.choices(*N_PERSON_MENTIONS)[0] if pool.persons else 0
            if n_sec or n_neg or n_pm:
                text, spans, negs = inject(text, lang, priors, n_sec, n_neg, pool, doc_index, cfg, n_pm)
        for s in spans + negs:
            v = text[s["start"]:s["end"]]
            assert len(v) == s["end"] - s["start"] and v.strip() == v and len(v) > 0, (s, v)
            s["value"] = v
        recs.append({"id": rec_id(i, ex), "source": "codeparrot/github-code-clean", "repo": ex["repo_name"], "license": ex["license"],
                     "language": lang, "path": ex["path"], "text": text, "spans": spans, "hard_negatives": negs,
                     "cluster_id": cid, "cluster_kind": kind, "cluster_size": n, "doc_index": doc_index, "entities": None})
    table = pool.table()
    recs[0]["entities"] = table
    for e in table:
        stats["entities"] += 1
        stats["entity_type"][e["entity_type"]] = stats["entity_type"].get(e["entity_type"], 0) + 1
        stats["docs_per_entity"][str(min(e["n_docs"], 6))] = stats["docs_per_entity"].get(str(min(e["n_docs"], 6)), 0) + 1
        stats["mentions_per_entity"][str(min(e["n_mentions"], 8))] = stats["mentions_per_entity"].get(str(min(e["n_mentions"], 8)), 0) + 1
        if e["owner_entity_id"]: stats["owned_entities"] += 1
    if n > 1:
        stats["recurring_entities"] += sum(1 for e in table if e["n_docs"] > 1)
        stats["multi_entities"] += len(table)
    stats["clusters"] += 1
    stats["cluster_kind"][kind] = stats["cluster_kind"].get(kind, 0) + 1
    stats["cluster_size"][str(n)] = stats["cluster_size"].get(str(n), 0) + 1
    for rec in recs:
        spans, negs, lang = rec["spans"], rec["hard_negatives"], rec["language"]
        line = json.dumps(rec, ensure_ascii=False) + "\n"
        fh.write(line)
        stats["files"] += 1; stats["bytes"] += len(line.encode("utf-8", "replace"))
        stats["injected_files" if spans else "clean_files"] += 1
        stats["spans"] += len(spans); stats["negs"] += len(negs)
        stats["lang"][lang] = stats["lang"].get(lang, 0) + 1
        stats["records_by_kind"][kind] = stats["records_by_kind"].get(kind, 0) + 1
        for s in spans:
            stats["label"][s["label"]] = stats["label"].get(s["label"], 0) + 1
            stats["value_type"][s["value_type"]] = stats["value_type"].get(s["value_type"], 0) + 1
            stats["context"][s["context"]] = stats["context"].get(s["context"], 0) + 1
            stats["surface_form"][s["surface_form_type"]] = stats["surface_form"].get(s["surface_form_type"], 0) + 1
        for s in negs:
            stats["value_type"]["neg:" + s["value_type"]] = stats["value_type"].get("neg:" + s["value_type"], 0) + 1

def process_shard(args):
    i, cfg = args
    R.seed(cfg["seed"] * 100003 + i); Faker.seed(cfg["seed"] * 100003 + i)
    priors = Priors(cfg["priors"])
    path, delete_after = fetch_shard(i, cfg["shard_dir"], cfg["keep_shards"])
    out_path = os.path.join(cfg["out"], f"part-{i:05d}.jsonl")
    stats = {"shard": i, "files": 0, "injected_files": 0, "clean_files": 0, "bytes": 0, "spans": 0, "negs": 0,
             "label": {}, "value_type": {}, "lang": {}, "context": {}, "skipped_license": 0, "skipped_size": 0,
             "clusters": 0, "cluster_kind": {}, "cluster_size": {}, "records_by_kind": {}, "entities": 0, "entity_type": {},
             "docs_per_entity": {}, "mentions_per_entity": {}, "owned_entities": 0, "recurring_entities": 0, "multi_entities": 0,
             "surface_form": {}, "peak_buffered_rows": 0}
    t0 = time.time()
    pf = pq.ParquetFile(path)
    plan, clusters = plan_clusters(pf, i, cfg, stats)
    stats["seconds_plan"] = round(time.time() - t0, 1)
    buffered = 0; idx = 0
    with open(out_path + ".tmp", "w", encoding="utf-8") as fh:
        for rg in range(pf.metadata.num_row_groups):
            for ex in pf.read_row_group(rg, columns=["code", "repo_name", "path", "language", "license"]).to_pylist():
                ck = plan.get(idx); idx += 1
                if ck is None:
                    continue
                c = clusters[ck]
                c.setdefault("buf", {})[idx - 1] = ex; buffered += 1
                stats["peak_buffered_rows"] = max(stats["peak_buffered_rows"], buffered)
                if len(c["buf"]) == len(c["members"]):
                    emit_cluster(c, i, cfg, priors, fh, stats)
                    buffered -= len(c["members"]); del clusters[ck]
    assert not clusters, f"{len(clusters)} clusters never completed"
    os.replace(out_path + ".tmp", out_path)
    if delete_after:
        os.remove(path)
    stats["seconds"] = round(time.time() - t0, 1)
    with open(os.path.join(cfg["out"], f"part-{i:05d}.stats.json"), "w") as fh:
        json.dump(stats, fh)
    print(f"shard {i}: {stats['files']} files, {stats['bytes']/1e6:.0f} MB, {stats['spans']} spans in {stats['seconds']}s ({stats['bytes']/1e6/stats['seconds']:.1f} MB/s)", flush=True)
    return stats

def main():
    ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
    ap.add_argument("--out", required=True)
    ap.add_argument("--priors", default=os.path.join(os.path.dirname(os.path.abspath(__file__)), "priors.json"))
    ap.add_argument("--target-bytes", type=float, default=2e9, help="stop launching new shards once output exceeds this (approx; one shard ~= 500-600 MB output)")
    ap.add_argument("--start-shard", type=int, default=0)
    ap.add_argument("--max-shards", type=int, default=N_SHARDS)
    ap.add_argument("--shard-dir", default="./shards", help="where parquet shards are cached/downloaded")
    ap.add_argument("--keep-shards", action="store_true", help="do not delete downloaded shards after processing")
    ap.add_argument("--workers", type=int, default=max(1, os.cpu_count() - 1))
    ap.add_argument("--inject-frac", type=float, default=0.75, help="fraction of files that get any injection (rest are pure negatives)")
    ap.add_argument("--min-bytes", type=int, default=200)
    ap.add_argument("--max-bytes", type=int, default=200_000)
    ap.add_argument("--seed", type=int, default=1)
    ap.add_argument("--max-rows", type=int, default=0, help="debug: only consider the first N rows of each shard")
    g = ap.add_argument_group("clusters / coreference (v3)")
    g.add_argument("--cluster-cap", type=int, default=8, help="max files per cluster; bigger repos are split into near-equal chunks")
    g.add_argument("--singleton-frac", type=float, default=0.35, help="share of files outside multi-file repos that stay singletons (rest go into bundles)")
    g.add_argument("--p-reuse", type=float, default=0.7, help="P(a secret slot reuses an existing cluster entity not yet in the file; its label replaces the drawn one)")
    g.add_argument("--p-repeat", type=float, default=0.15, help="P(a secret slot repeats an entity already mentioned in the same file)")
    g.add_argument("--p-owner", type=float, default=0.5, help="P(a new secret is owned by a cluster person, when persons exist)")
    g.add_argument("--no-persons", action="store_true", help="v2-compatible label set: no person entities / person_name / username spans")
    g.add_argument("--no-bundles", action="store_true", help="no synthetic bundles (all loose files stay singletons)")
    g.add_argument("--no-repo-clusters", action="store_true", help="treat multi-file repos like loose files")
    a = ap.parse_args()
    os.makedirs(a.out, exist_ok=True)
    cfg = dict(priors=a.priors, out=a.out, shard_dir=a.shard_dir, keep_shards=a.keep_shards, inject_frac=a.inject_frac, min_bytes=a.min_bytes, max_bytes=a.max_bytes, seed=a.seed, max_rows=a.max_rows,
               cluster_cap=a.cluster_cap, singleton_frac=1.0 if a.no_bundles else a.singleton_frac, p_reuse=a.p_reuse, p_repeat=a.p_repeat, p_owner=a.p_owner,
               no_persons=a.no_persons, no_repo_clusters=a.no_repo_clusters)
    t0 = time.time(); total = 0; done = []
    shard_iter = iter(range(a.start_shard, min(N_SHARDS, a.start_shard + a.max_shards)))
    with Pool(a.workers) as pool:
        pending = []
        def launch():
            i = next(shard_iter, None)
            if i is not None:
                pending.append(pool.apply_async(process_shard, ((i, cfg),)))
        for _ in range(a.workers):
            launch()
        while pending:
            r = pending.pop(0)
            st = r.get(); done.append(st); total += st["bytes"]
            if total < a.target_bytes:
                launch()
    dt = time.time() - t0
    summary = {"shards": len(done), "files": sum(s["files"] for s in done), "bytes": total, "spans": sum(s["spans"] for s in done),
               "hard_negatives": sum(s["negs"] for s in done), "clusters": sum(s["clusters"] for s in done), "entities": sum(s["entities"] for s in done),
               "seconds": round(dt, 1), "MB_per_s": round(total / 1e6 / dt, 2), "workers": a.workers,
               "config": {k: cfg[k] for k in ("seed", "inject_frac", "cluster_cap", "singleton_frac", "p_reuse", "p_repeat", "p_owner", "no_persons")}}
    with open(os.path.join(a.out, "summary.json"), "w") as fh:
        json.dump(summary, fh, indent=1)
    print(json.dumps(summary, indent=1))

if __name__ == "__main__":
    main()