tools/build_eval.py

"""Build the three measurement corpora.

    python tools/build_eval.py --build builds/muse-glimmer-30b \
        --raw-eval /workspace/calib-build/raw_eval \
        --wiki /workspace/calib-build/wiki_cache.jsonl \
        --wikitext /workspace/calib-build/wikitext

Three corpora, each sized so that llama-perplexity at ctx 4096 scores a useful
number of positions.  That tool evaluates the second half of every context
window, so a corpus of N tokens yields floor(N/4096) chunks and half that many
times 4096 scored positions -- the corpus has to be twice the size of the
measurement you actually want.

  eval/neutral   prose, no code, many scripts
  eval/code      source code; the previous eval_neutral.txt renamed and kept
                 byte-identical for comparability, plus an extension
  eval/agentic   conversations in the model's own markup

Selection is by exclusion.  A document enters a corpus only if it shares no
13-gram with the calibration files, with wikitext, or with the eval corpora
already built.  Ordering is fixed and the whole thing is seeded, so a rerun
produces the same bytes.

One consequence worth stating plainly: llama-perplexity has no --parse-special
(the flag is registered for LLAMA_EXAMPLE_IMATRIX only).  The special tokens
eval/agentic is required to contain will therefore be scored as their literal
characters, which is a different measurement from what the model sees at
inference.  eval/neutral and eval/code are asserted free of those strings so
they are unaffected.
"""

from __future__ import annotations

import argparse
import copy
import glob
import json
import os
import random
import sys
import unicodedata
from collections import Counter, defaultdict

import numpy as np

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

import glimmer_fmt as G
import poollib as P

SEED = 20260810
NGRAM = 13
CTX = 4096
TARGET_LO, TARGET_HI = 350_000, 400_000
DOC_SEP = "\n\n"
SPECIALS = ("<|start|>", "<|message|>", "<|eot|>", "<|eom|>", "<|patch|>",
            "<|begin_of_text|>", "<|end_of_text|>", "<|video|>")
MAX_DUP_LINE_PERCENT = 2.0

# Repositories reserved for measurement only. None of these appear in
# pipeline/sources.py REPOS, so nothing here can be in the pool; the four
# originals were not enough material once overlap filtering had run.
EXTRA_EVAL_REPOS = {
    "got":         ("https://github.com/sindresorhus/got",   "MIT"),
    "jinja":       ("https://github.com/pallets/jinja",      "BSD-3-Clause"),
    "rust-regex":  ("https://github.com/rust-lang/regex",    "MIT OR Apache-2.0"),
    "googletest":  ("https://github.com/google/googletest",  "BSD-3-Clause"),
    "black":       ("https://github.com/psf/black",          "MIT"),
    "express":     ("https://github.com/expressjs/express",  "MIT"),
    "tokio-bytes": ("https://github.com/tokio-rs/bytes",     "MIT"),
}

# Unicode ranges, enough to classify a corpus by writing system.
SCRIPTS = [
    ("Latin",     [(0x0041, 0x024F), (0x1E00, 0x1EFF)]),
    ("Cyrillic",  [(0x0400, 0x04FF), (0x0500, 0x052F)]),
    ("Greek",     [(0x0370, 0x03FF), (0x1F00, 0x1FFF)]),
    ("Arabic",    [(0x0600, 0x06FF), (0x0750, 0x077F), (0xFB50, 0xFDFF)]),
    ("Hebrew",    [(0x0590, 0x05FF)]),
    ("Armenian",  [(0x0530, 0x058F)]),
    ("Georgian",  [(0x10A0, 0x10FF), (0x1C90, 0x1CBF)]),
    ("Devanagari", [(0x0900, 0x097F)]),
    ("Bengali",   [(0x0980, 0x09FF)]),
    ("Tamil",     [(0x0B80, 0x0BFF)]),
    ("Thai",      [(0x0E00, 0x0E7F)]),
    ("Myanmar",   [(0x1000, 0x109F)]),
    ("Ethiopic",  [(0x1200, 0x137F)]),
    ("Han",       [(0x4E00, 0x9FFF), (0x3400, 0x4DBF), (0xF900, 0xFAFF)]),
    ("Hiragana",  [(0x3040, 0x309F)]),
    ("Katakana",  [(0x30A0, 0x30FF)]),
    ("Hangul",    [(0xAC00, 0xD7AF), (0x1100, 0x11FF), (0x3130, 0x318F)]),
]


def script_of(ch: str) -> str | None:
    cp = ord(ch)
    if not ch.isalpha():
        return None
    for name, ranges in SCRIPTS:
        for lo, hi in ranges:
            if lo <= cp <= hi:
                return name
    return "Other"


def script_mix(text: str) -> dict:
    c = Counter()
    for ch in text:
        s = script_of(ch)
        if s:
            c[s] += 1
    tot = sum(c.values()) or 1
    return {k: round(100.0 * v / tot, 2) for k, v in c.most_common()}


def duplicate_line_share(text: str, min_len: int = 0,
                         ignore_lines: set | None = None) -> tuple[float, list]:
    lines = [ln for ln in text.split("\n") if len(ln.strip()) >= min_len and ln.strip()]
    if ignore_lines:
        lines = [ln for ln in lines if ln not in ignore_lines]
    if not lines:
        return 0.0, []
    c = Counter(lines)
    dup = sum(n - 1 for n in c.values() if n > 1)
    return 100.0 * dup / len(lines), c.most_common(8)


class Excluder:
    """Everything a new eval document must not overlap.

    Two tiers on purpose.  The static side holds wikitext plus the calibration
    files -- about a hundred million 13-grams -- and is sorted once.  The
    dynamic side holds what has been accepted so far and grows one document at
    a time; keeping it a plain set avoids re-sorting the hundred million on
    every acceptance, which is the difference between seconds and hours.
    """

    def __init__(self, ignore: set[int] | None = None) -> None:
        self.static = P.ShingleIndex(NGRAM)
        self.dynamic: set[int] = set()
        self.sources: list[str] = []
        # The chat template emits a fixed preamble, a worked example and a
        # recipients line into every conversation that declares tools. Those
        # 13-grams are the template's, not the corpus's, and they appear
        # identically in calibration and in measurement. Counting them as
        # overlap rejected 100% of eval/agentic on the first run.
        self.ignore = ignore or set()

    def add_static(self, text: str | None, label: str,
                   hashes: np.ndarray | None = None) -> None:
        if hashes is not None:
            self.static.add_hashes(hashes)
        else:
            self.static.add(text)
        self.sources.append(label)

    def freeze(self) -> "Excluder":
        self.static.finalise()
        return self

    def add_text(self, text: str, label: str) -> None:
        self.dynamic.update(int(x) for x in P.shingles(text, NGRAM)
                            if int(x) not in self.ignore)

    def overlap(self, text: str) -> int:
        h = P.shingles(text, NGRAM)
        if h.size == 0:
            return 0
        if self.ignore:
            h = np.array([x for x in h.tolist() if x not in self.ignore],
                         dtype=np.uint64)
            if h.size == 0:
                return 0
        n = int(self.static.contains(h).sum())
        if n:
            return n
        if self.dynamic:
            n += sum(1 for x in h.tolist() if x in self.dynamic)
        return n


def take_until(cands, excl: Excluder, tok, lo: int, hi: int, label: str,
               max_dup_percent: float | None = None,
               allow_intra_repeat: bool = False):
    """Greedily accept documents that overlap nothing already accepted.

    `max_dup_percent` additionally rejects a document whose lines are mostly
    already in the corpus, which is how the duplicate-line ceiling is met
    without editing anybody's text.

    `allow_intra_repeat` says the zero-overlap rule applies only *between*
    corpora, not inside one. Generated conversations necessarily repeat their
    own framing, and enforcing zero within a corpus rejected 613 of 615
    candidates; repetition inside a corpus is what the duplicate-line ceiling
    is for.
    """
    out, total, rejected, dup_rejected = [], 0, 0, 0
    seen_lines: Counter = Counter()
    n_lines = n_dup = 0
    for d in cands:
        if total >= lo:
            break
        text = d["text"]
        if excl.overlap(text):
            rejected += 1
            continue
        n = tok.count(text)
        if n == 0 or total + n > hi:
            continue
        if max_dup_percent is not None:
            lines = [ln for ln in text.split("\n") if ln.strip()]
            add_dup = sum(1 for ln in lines if seen_lines[ln] > 0)
            if n_lines + len(lines) and \
                    100.0 * (n_dup + add_dup) / (n_lines + len(lines)) > max_dup_percent:
                dup_rejected += 1
                continue
            seen_lines.update(lines)
            n_lines += len(lines)
            n_dup += add_dup
        out.append((d, text, n))
        total += n
        if not allow_intra_repeat:
            excl.add_text(text, label)      # later docs must not repeat it either
    if allow_intra_repeat:                  # seal the corpus against the next one
        for _, text, _ in out:
            excl.add_text(text, label)
    extra = f", {dup_rejected} rejected for duplicate lines" if dup_rejected else ""
    print(f"    {len(out)} documents accepted, {rejected} rejected for overlap"
          f"{extra}, {total:,} tokens")
    return out, total


def write_corpus(path: str, docs, manifest_path: str, extra_head: str = "") -> str:
    body = extra_head + DOC_SEP.join(t for _, t, _ in docs) + "\n"
    os.makedirs(os.path.dirname(path), exist_ok=True)
    with open(path, "w", encoding="utf-8") as f:
        f.write(body)
    with open(manifest_path, "w", encoding="utf-8") as f:
        for d, text, n in docs:
            f.write(json.dumps({
                "id": d.get("id", ""), "source": d["source"], "path": d["path"],
                "license": d.get("license", ""), "lang": d.get("lang", ""),
                "origin": d.get("origin", ""), "tokens": n, "chars": len(text),
            }, ensure_ascii=False) + "\n")
    return body


def measure(name: str, path: str, tok, must_have_specials: bool,
            template_lines: set | None = None) -> dict:
    text = open(path, encoding="utf-8").read()
    ids = tok.encode(text)
    chunks = len(ids) // CTX
    scored = chunks * (CTX // 2)
    c = Counter(ids)
    dup, top = duplicate_line_share(text)
    dup_sig, _ = duplicate_line_share(text, min_len=12)
    dup_notmpl, _ = duplicate_line_share(text, ignore_lines=template_lines)
    present = [s for s in SPECIALS if s in text]
    ok_specials = bool(present) if must_have_specials else not present
    out = {
        "corpus": name, "path": os.path.relpath(path, P.REPO_ROOT),
        "tokens": len(ids), "chars": len(text),
        "chunks_at_ctx_4096": chunks, "scored_positions": scored,
        "vocab_rows": tok.n_vocab,
        "vocab_covered": len(c),
        "vocab_covered_percent": round(100.0 * len(c) / tok.n_vocab, 3),
        "duplicate_line_percent": round(dup, 2),
        "duplicate_line_percent_len_ge_12": round(dup_sig, 2),
        "duplicate_line_percent_excluding_chat_template": round(dup_notmpl, 2),
        "top_repeated_lines": [[l, n] for l, n in top],
        "script_mix_percent": script_mix(text),
        "special_tokens_present": present,
        "special_token_rule_ok": ok_specials,
    }
    print(f"\n--- {name} ---")
    print(f"  {len(ids):,} tokens, {chunks} chunks at ctx {CTX}, "
          f"{scored:,} scored positions")
    print(f"  vocabulary: {len(c):,} of {tok.n_vocab:,} rows "
          f"({100.0*len(c)/tok.n_vocab:.2f}%)")
    print(f"  duplicate lines: {dup:.2f}%  (>=12 chars: {dup_sig:.2f}%"
          f"{f', excl. chat template: {dup_notmpl:.2f}%' if template_lines else ''})")
    print(f"  scripts: {json.dumps(out['script_mix_percent'])[:150]}")
    print(f"  special tokens {'present' if present else 'absent'} "
          f"-> rule {'ok' if ok_specials else 'VIOLATED'}")
    return out


def main() -> int:
    ap = argparse.ArgumentParser()
    ap.add_argument("--build", required=True)
    ap.add_argument("--raw-eval", required=True)
    ap.add_argument("--wiki", required=True)
    ap.add_argument("--wikitext", required=True)
    ap.add_argument("--tokenizer", required=True)
    ap.add_argument("--repo", default=P.REPO_ROOT)
    args = ap.parse_args()

    rng = random.Random(SEED)
    tok = P.TargetTokenizer(args.tokenizer)
    eval_root = os.path.join(args.repo, "eval")

    # ------------------------------------------------- step 0: rename in place
    legacy_src = os.path.join(args.repo, "eval_neutral.txt")
    legacy_dst = os.path.join(eval_root, "code", "eval_code.txt")
    os.makedirs(os.path.dirname(legacy_dst), exist_ok=True)
    if os.path.exists(legacy_src) and not os.path.exists(legacy_dst):
        with open(legacy_src, encoding="utf-8") as f:
            legacy_text = f.read()
        with open(legacy_dst, "w", encoding="utf-8") as f:
            f.write(legacy_text)
        print(f"eval_neutral.txt -> eval/code/eval_code.txt "
              f"({len(legacy_text):,} chars, byte-identical)")
    else:
        legacy_text = open(legacy_dst, encoding="utf-8").read()
    print(f"  legacy corpus: {tok.count(legacy_text):,} tokens under this tokenizer")

    # ---------------------------------------------- the exclusion set, layered
    print("\nbuilding exclusion index")
    dummy = [{"type": "function", "function": {
        "name": "ns.fn", "description": "d",
        "parameters": {"type": "object", "properties": {}}}}]
    # Every reasoning strength, with and without a tool block. Building this
    # for "high" alone silently pinned the corpus to "high": the 13-grams
    # around `Reasoning strength: low.` were then *not* exempt, matched
    # calib_train (which uses all four levels), and every low/medium/xhigh
    # conversation was rejected as an overlap.
    scaffolds = [G.render([{"role": "user", "content": ""},
                           {"role": "assistant", "content": ""}],
                          tools=t, reasoning_strength=rs,
                          knowledge_cutoff="2026-01-04", current_date="2026-08-10")
                 for rs in G.REASONING_STRENGTHS for t in (dummy, None)]
    scaffold = scaffolds[0]
    template_grams = {int(x) for sc in scaffolds for x in P.shingles(sc, NGRAM)}
    print(f"  {len(template_grams)} {NGRAM}-grams belong to the chat template "
          f"itself and are excluded from every comparison")
    forbidden = Excluder(ignore=template_grams)
    for f in ("calib_train.txt", "calib_longctx.txt"):
        p = os.path.join(args.build, f)
        if os.path.exists(p):
            forbidden.add_static(open(p, encoding="utf-8").read(), f)
            print(f"  + {f}")
    import pyarrow.parquet as pq
    wt_parts = []
    for f in sorted(glob.glob(os.path.join(args.wikitext, "**", "*.parquet"),
                              recursive=True)):
        wt_parts.append(P.shingles("\n".join(pq.read_table(f).column("text").to_pylist()),
                                   NGRAM))
    forbidden.add_static(None, "wikitext-103-raw-v1",
                         hashes=np.unique(np.concatenate(wt_parts)))
    del wt_parts
    print("  + wikitext-103-raw-v1")

    import sources as S
    results = []

    # ------------------------------------------------------------ eval/code
    # The legacy file stays exactly as it is, so it is added to the exclusion
    # set rather than filtered: the extension must not repeat it.
    print("\n[eval/code] extending")
    forbidden.add_static(legacy_text, "eval/code/eval_code.txt")
    forbidden.freeze()
    print(f"  {len(forbidden.static):,} distinct {NGRAM}-grams in the static "
          f"exclusion set ({', '.join(forbidden.sources)})")
    legacy_paths = {(r["source"], r["path"])
                    for r in P.read_jsonl(os.path.join(args.repo,
                                                       "eval_neutral.manifest.jsonl"))}
    code_docs, agentic_pool = [], []
    all_eval_repos = {r: (u, l) for r, (u, l, _) in S.EVAL_REPOS.items()}
    all_eval_repos.update(EXTRA_EVAL_REPOS)
    # The legacy eval_code.txt was drawn from the four original repositories,
    # so those stay on the code side; the grounding set gets repositories that
    # appear in no other file at all.
    GROUNDING_REPOS = {"rust-regex", "googletest", "black", "tokio-bytes"}
    for repo in sorted(all_eval_repos):
        url, lic = all_eval_repos[repo]
        root = os.path.join(args.raw_eval, repo)
        if not os.path.isdir(root):
            print(f"  ! missing eval clone: {repo}")
            continue
        for rel, lang, text in sorted(S.walk_repo(root, repo), key=lambda t: t[0]):
            if (repo, rel) in legacy_paths:
                continue
            d = {"id": P.make_id(repo, rel, text), "source": repo, "path": rel,
                 "license": lic, "lang": lang, "origin": "repo_file", "text": text,
                 "upstream": url}
            (agentic_pool if repo in GROUNDING_REPOS else code_docs).append(d)
    code_docs.sort(key=lambda d: (d["source"], d["path"]))
    rng.shuffle(code_docs)
    need_lo = max(0, TARGET_LO - tok.count(legacy_text))
    need_hi = max(0, TARGET_HI - tok.count(legacy_text))
    ext, ext_tok = take_until(code_docs, forbidden, tok, need_lo, need_hi, "eval/code")
    write_corpus(os.path.join(eval_root, "code", "eval_code_ext.txt"), ext,
                 os.path.join(eval_root, "code", "eval_code_ext.manifest.jsonl"))
    with open(os.path.join(eval_root, "code", "eval_code_full.txt"), "w",
              encoding="utf-8") as f:
        f.write(legacy_text.rstrip("\n") + DOC_SEP +
                DOC_SEP.join(t for _, t, _ in ext) + "\n")
    results.append(measure("eval/code", os.path.join(eval_root, "code",
                                                     "eval_code_full.txt"), tok, False))

    # --------------------------------------------------------- eval/neutral
    print("\n[eval/neutral] multilingual prose")
    wiki = [json.loads(l) for l in open(args.wiki, encoding="utf-8")]
    by_lang = defaultdict(list)
    for r in wiki:
        by_lang[r["lang"]].append(r)
    # Interleave languages so that a Latin-heavy language cannot fill the
    # corpus before the others are reached; non-Latin first for the same reason.
    NONLATIN = {"zh", "ja", "ko", "ru", "uk", "ar", "fa", "he", "el", "hi", "bn",
                "ta", "ka", "hy", "my", "am", "th"}
    order = sorted(by_lang, key=lambda l: (l not in NONLATIN, l))
    for l in order:
        by_lang[l].sort(key=lambda r: r["id"])
    interleaved = []
    i = 0
    while any(by_lang[l][i:i + 1] for l in order):
        for l in order:
            if i < len(by_lang[l]):
                r = by_lang[l][i]
                interleaved.append({"id": r["id"], "source": r["source"],
                                    "path": r["path"], "license": r.get("license", "CC-BY-SA-4.0"),
                                    "lang": r["lang"], "origin": "wikipedia",
                                    "text": r["text"]})
        i += 1
    neutral, n_tok = take_until(interleaved, forbidden, tok, TARGET_LO, TARGET_HI,
                                "eval/neutral", max_dup_percent=MAX_DUP_LINE_PERCENT)
    write_corpus(os.path.join(eval_root, "neutral", "eval_neutral.txt"), neutral,
                 os.path.join(eval_root, "neutral", "eval_neutral.manifest.jsonl"))
    r = measure("eval/neutral", os.path.join(eval_root, "neutral", "eval_neutral.txt"),
                tok, False)
    r["languages"] = len({d["lang"] for d, _, _ in neutral})
    print(f"  languages: {r['languages']}")
    results.append(r)

    # --------------------------------------------------------- eval/agentic
    print("\n[eval/agentic] native markup")
    import gen_agentic_eval as GAE
    convs = GAE.build(120, SEED + 7)
    convs += GAE.build_grounded(agentic_pool, min(len(agentic_pool), 900), SEED + 11)
    print(f"  {len(convs)} candidate conversations "
          f"({len(agentic_pool)} held-out files available for grounding)")
    ag_docs = []
    for c in convs:
        text = G.render(c["messages"], tools=c["tools"],
                        reasoning_strength=c["reasoning_strength"],
                        knowledge_cutoff="2026-01-04", current_date="2026-08-10",
                        namespace_descriptions=GAE.NS_DESCRIPTIONS, add_bos=True)
        ag_docs.append({
            "id": P.make_id("eval-agentic", f"{c['scenario']}/{c['index']}", text),
            "source": f"synthetic/eval-agentic:{c['scenario']}",
            "path": f"eval_agentic/{c['scenario']}/{c['index']:04d}",
            "license": "CC0-1.0 (generated; quoted file excerpts keep their upstream licence)",
            "lang": "chat", "origin": "synth_agentic_eval", "text": text,
            "grounded_in": c.get("grounded_in", ""),
            "reasoning_strength": c["reasoning_strength"]})
    agentic, a_tok = take_until(ag_docs, forbidden, tok, TARGET_LO, TARGET_HI,
                                "eval/agentic", allow_intra_repeat=True)
    write_corpus(os.path.join(eval_root, "agentic", "eval_agentic.txt"), agentic,
                 os.path.join(eval_root, "agentic", "eval_agentic.manifest.jsonl"))
    tmpl_lines = {ln for sc in scaffolds for ln in sc.split("\n") if ln.strip()}
    r = measure("eval/agentic",
                os.path.join(eval_root, "agentic", "eval_agentic.txt"), tok, True,
                template_lines=tmpl_lines)
    r["reasoning_strengths"] = dict(Counter(d.get("reasoning_strength", "")
                                            for d, _, _ in agentic))
    print(f"  reasoning strengths: {r['reasoning_strengths']}")
    results.append(r)

    # ------------------------------------------------------- pairwise checks
    print("\n[cross-checks] 13-gram intersections")
    files = {
        "calib_train": os.path.join(args.build, "calib_train.txt"),
        "calib_longctx": os.path.join(args.build, "calib_longctx.txt"),
        "eval/code (full)": os.path.join(eval_root, "code", "eval_code_full.txt"),
        "eval/code (extension only)": os.path.join(eval_root, "code", "eval_code_ext.txt"),
        "eval/neutral": os.path.join(eval_root, "neutral", "eval_neutral.txt"),
        "eval/agentic": os.path.join(eval_root, "agentic", "eval_agentic.txt"),
    }
    sh = {k: np.unique(P.shingles(open(v, encoding="utf-8").read(), NGRAM))
          for k, v in files.items() if os.path.exists(v)}
    matrix = {}
    names = list(sh)
    # eval/code (full) contains the extension by construction, and both
    # calibration files are drawn from one pool; neither pair is a violation.
    contained = {("eval/code (full)", "eval/code (extension only)"),
                 ("calib_train", "calib_longctx")}
    ignore = template_grams
    for i, a in enumerate(names):
        for b in names[i + 1:]:
            inter = np.intersect1d(sh[a], sh[b], assume_unique=True)
            n = int(inter.size)
            n_adj = int(sum(1 for x in inter.tolist() if x not in ignore))
            matrix[f"{a} ^ {b}"] = {"shared": n, "excluding_chat_template": n_adj}
            if (a, b) in contained or (b, a) in contained:
                note = "   (containment, expected)"
            elif n_adj:
                note = "   <-- NON-ZERO"
            else:
                note = ""
            print(f"  {a:28} ^ {b:28} {n:7} raw {n_adj:7} adj{note}")
    out = {"corpora": results, "intersections": matrix, "ngram": NGRAM,
           "ctx": CTX, "tokenizer": args.tokenizer}
    with open(os.path.join(eval_root, "eval-report.json"), "w", encoding="utf-8") as f:
        json.dump(out, f, indent=2, ensure_ascii=False)
        f.write("\n")
    print(f"\nwrote {os.path.join(eval_root, 'eval-report.json')}")
    return 0


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