| """Pairwise 13-gram intersections between calibration and measurement files. |
| |
| python tools/crosscheck.py builds/muse-glimmer-30b/calib_train.txt \ |
| builds/muse-glimmer-30b/calib_longctx.txt eval/*/*.txt |
| |
| Reports each pair twice: the raw count, and the count with the chat template's |
| own scaffolding removed. The template emits a fixed preamble, a worked example |
| and a `Reasoning strength:` line into every conversation that declares tools; |
| those 13-grams belong to the model, not to either corpus, and counting them |
| makes an agentic measurement corpus look contaminated by an agentic |
| calibration corpus when the only thing they share is the format. |
| """ |
|
|
| from __future__ import annotations |
|
|
| import argparse |
| import json |
| import os |
| import sys |
|
|
| import numpy as np |
|
|
| HERE = os.path.dirname(os.path.abspath(__file__)) |
| sys.path.insert(0, HERE) |
|
|
| import glimmer_fmt as G |
| import poollib as P |
|
|
| NGRAM = 13 |
|
|
|
|
| def template_grams() -> set[int]: |
| dummy = [{"type": "function", "function": { |
| "name": "ns.fn", "description": "d", |
| "parameters": {"type": "object", "properties": {}}}}] |
| out: set[int] = set() |
| rendered = [] |
| for rs in G.REASONING_STRENGTHS: |
| for t in (dummy, None): |
| for msgs in ([{"role": "user", "content": ""}, |
| {"role": "assistant", "content": ""}], |
| [{"role": "system", "content": ""}, |
| {"role": "user", "content": ""}]): |
| s = G.render(msgs, tools=t, reasoning_strength=rs, |
| knowledge_cutoff="2026-01-04", |
| current_date="2026-08-10") |
| rendered.append(s) |
| out |= {int(x) for x in P.shingles(s, NGRAM)} |
| |
| |
| |
| |
| for a in rendered: |
| for b in rendered: |
| out |= {int(x) for x in P.shingles(a + "\n\n" + b, NGRAM)} |
| return out |
|
|
|
|
| def main() -> int: |
| ap = argparse.ArgumentParser() |
| ap.add_argument("files", nargs="+") |
| ap.add_argument("--json-out") |
| args = ap.parse_args() |
|
|
| ig = template_grams() |
| print(f"{len(ig)} {NGRAM}-grams belong to the chat template itself\n") |
| sh = {} |
| for f in args.files: |
| if not os.path.exists(f): |
| continue |
| sh[os.path.relpath(f, P.REPO_ROOT)] = np.unique( |
| P.shingles(open(f, encoding="utf-8").read(), NGRAM)) |
|
|
| names, matrix, bad = list(sh), {}, 0 |
| for i, a in enumerate(names): |
| for b in names[i + 1:]: |
| inter = np.intersect1d(sh[a], sh[b], assume_unique=True) |
| adj = int(sum(1 for x in inter.tolist() if x not in ig)) |
| matrix[f"{a} ^ {b}"] = {"shared": int(inter.size), |
| "excluding_chat_template": adj} |
| note = "" |
| if adj: |
| |
| if not ("calib" in a and "calib" in b) and \ |
| not (a.startswith(os.path.dirname(b)) or b.startswith(os.path.dirname(a))): |
| note = " <-- NON-ZERO" |
| bad += 1 |
| else: |
| note = " (same side, expected)" |
| print(f" {a:44} ^ {b:44} {inter.size:7} raw {adj:7} adj{note}") |
| print(f"\n{bad} calibration/measurement pairs still intersect") |
| if args.json_out: |
| with open(args.json_out, "w") as f: |
| json.dump(matrix, f, indent=2) |
| f.write("\n") |
| return 0 |
|
|
|
|
| if __name__ == "__main__": |
| sys.exit(main()) |
|
|
