CodeEditSearchTrain

A code-edit retrieval training set, derived from bigcode/commitpackft and decontaminated against the cassanof/CodeEditSearch eval.

Each example contains an instruction (the commit message), a unified diff, and the before/after file contents, plus the source commit SHA.

How the dataset is built

We start from bigcode/commitpackft, a large collection of code-edit commits across many programming languages, and keep every language with at least 1,000 commits — 47 languages in total. For each commit we keep the message as the natural-language instruction, render a short unified diff between the file before and after the change, and store the full file contents on both sides. Commits that don't have both a before and an after file, or that have an empty message, are discarded.

The harder problem is making sure none of the kept commits leak into the cassanof/CodeEditSearch evaluation set. We apply three filters, each catching something the previous one would miss:

1. Same commit. If a commit's SHA already appears in the eval, we drop it, because same git commit means same code change.
2. Same text after normalization. Two commits can describe the same change with different SHAs (cherry-picks, rebases, mirrored repos). We lowercase, normalize Unicode and collapse whitespace, then hash every text field with xxHash-64 and compare against the same hashes computed over the eval. Anything that collides is dropped.
3. Near-duplicate text. Even after normalization, paraphrased or partially edited versions slip through. We extract every 13-word sequence from each field and compute what fraction of them also appear in the eval (containment). If any field crosses 50% overlap, the commit is dropped.

Steps 2 and 3 follow the decontamination methodology described in our English models blog post ("Decontaminated BEIR" section) and our multilingual models blog post.

The eval only covers 13 languages, but the contamination index pools everything from those 13 into one big set, and we check every commit in all 47 source languages against it, so a Python snippet that also appears verbatim in a YAML commit message would still be caught.

End result: 678,295 source commits → 599,698 with both file contents → 578,203 after dropping shared SHAs → 573,593 after the exact-text filter → 560,922 kept after the near-duplicate filter. Per-language numbers are in the report below.

Generation code

Python code to build and decontaminate the dataset

"""
Build a code-edit retrieval training set from `bigcode/commitpackft`,
deduped against `cassanof/CodeEditSearch` via (A) commit-SHA exclusion,
(B) normalized xxHash-64, and (C) 13-gram containment.
"""

from __future__ import annotations

import argparse
import difflib
import json
import logging
import os
import re
import sys
import unicodedata
import urllib.request
from pathlib import Path

import xxhash
from datasets import load_dataset, Dataset
from dotenv import load_dotenv
from tqdm import tqdm

logger = logging.getLogger(__name__)


# 13 languages covered by the cassanof/CodeEditSearch eval:
EVAL_LANGUAGES = [
    "c",
    "c++",
    "go",
    "java",
    "javascript",
    "php",
    "python",
    "ruby",
    "rust",
    "scala",
    "shell",
    "swift",
    "typescript",
]

SHUFFLE_SEED = 42
EVAL_DATASET = "cassanof/CodeEditSearch"
SOURCE_DATASET = "bigcode/commitpackft"
SIZE_API = "https://datasets-server.huggingface.co/size?dataset={dataset}"

REPORT_DIR = Path(__file__).resolve().parent / "output"

WHITESPACE_RE = re.compile(r"\s+")


def discover_languages(dataset: str, min_rows: int) -> list[str]:
    """Fetch all configs from HF datasets-server with at least `min_rows` rows."""
    url = SIZE_API.format(dataset=dataset)
    with urllib.request.urlopen(url) as resp:
        data = json.load(resp)
    configs = data["size"]["configs"]
    kept = [c["config"] for c in configs if c["num_rows"] >= min_rows]
    logger.info(
        "discovered %d/%d configs in %s with >= %d rows",
        len(kept), len(configs), dataset, min_rows,
    )
    return sorted(kept)


def collect_eval_shas(languages: list[str]) -> set[str]:
    """Pool every commit SHA appearing in the eval dataset across the given languages."""
    shas: set[str] = set()
    for lang in languages:
        try:
            ds = load_dataset(EVAL_DATASET, lang, split="train")
        except Exception as e:
            logger.warning("could not load eval for lang=%s: %s", lang, e)
            continue
        for row in ds:
            sha = row.get("commit")
            if sha:
                shas.add(sha)
    logger.info("collected %d eval commit SHAs across %d languages", len(shas), len(languages))
    return shas


def has_nonempty_contents(row: dict) -> bool:
    """True if both `old_contents` and `new_contents` are present and non-empty."""
    return bool(row.get("old_contents")) and bool(row.get("new_contents"))


def build_unified_diff(old: str, new: str, old_file: str = "a", new_file: str = "b") -> str:
    """Render a unified diff (3 lines of context) between two file contents."""
    return "".join(difflib.unified_diff(
        old.splitlines(keepends=True),
        new.splitlines(keepends=True),
        fromfile=old_file,
        tofile=new_file,
        n=3,
    ))


def normalize(text: str) -> str:
    """NFKD-fold, lowercase, and collapse whitespace runs to single spaces."""
    text = unicodedata.normalize("NFKD", text).lower()
    return WHITESPACE_RE.sub(" ", text).strip()


def hash64(text: str) -> int:
    """64-bit xxh64 of text bytes."""
    return xxhash.xxh64(text.encode("utf-8", errors="replace")).intdigest()


def word_ngrams(text: str, n: int = 13) -> set[str]:
    """Set of contiguous n-word phrases from whitespace-split text."""
    words = text.split()
    if len(words) < n:
        return set()
    return {" ".join(words[i:i + n]) for i in range(len(words) - n + 1)}


def field_hash(text: str) -> int | None:
    """Normalized hash of one field; None if the field is empty after normalization."""
    norm = normalize(text or "")
    if not norm:
        return None
    return hash64(norm)


def build_eval_overlap_index(languages: list[str], n: int = 13) -> tuple[set[int], set[str]]:
    """Index per-field normalized hashes and pooled per-field n-grams across all eval rows."""
    field_hashes: set[int] = set()
    ngram_set: set[str] = set()
    logger.info("building per-field contamination index over eval rows...")
    total = 0
    for lang in languages:
        try:
            ds = load_dataset(EVAL_DATASET, lang, split="train")
        except Exception:
            raise RuntimeError(f"failed to load eval dataset for lang={lang}; check that EVAL_LANGUAGES is correct and that the dataset is accessible")
        for row in tqdm(ds, desc=f"overlap-index[{lang}]"):
            fields = (
                (row.get("instruction") or "").strip(),
                row.get("diff") or "",
                row.get("before") or "",
                row.get("after") or "",
            )
            if not any(fields):
                continue
            for f in fields:
                h = field_hash(f)
                if h is not None:
                    field_hashes.add(h)
                ngram_set.update(word_ngrams(normalize(f), n=n))
            total += 1
    logger.info(
        "indexed %d eval rows: %d field-hashes, %d %d-grams",
        total, len(field_hashes), len(ngram_set), n,
    )
    return field_hashes, ngram_set


def contamination_reason(
    instruction: str,
    diff: str,
    old_contents: str,
    new_contents: str,
    eval_field_hashes: set[int],
    eval_ngrams: set[str],
    n: int = 13,
    threshold: float = 0.5,
) -> str:
    """Return "hash"/"ngram" if any of the 4 fields matches eval, else "" (clean)."""
    fields = (instruction, diff, old_contents, new_contents)
    for f in fields:
        h = field_hash(f)
        if h is not None and h in eval_field_hashes:
            return "hash"
    for f in fields:
        sample_ngrams = word_ngrams(normalize(f), n=n)
        if not sample_ngrams:
            continue
        if len(sample_ngrams & eval_ngrams) / len(sample_ngrams) >= threshold:
            return "ngram"
    return ""


def build_language(
    lang: str,
    eval_shas: set[str],
    eval_field_hashes: set[int],
    eval_ngrams: set[str],
    output_repo: str,
    max_per_lang: int | None,
    ngram_n: int = 13,
    overlap_threshold: float = 0.5,
) -> dict:
    """Filter one language of the source dataset against the eval, push to the Hub, and return its funnel."""
    logger.info("lang=%s loading source from %s", lang, SOURCE_DATASET)
    ds = load_dataset(
        "json",
        data_files=f"hf://datasets/{SOURCE_DATASET}/data/{lang}/data.jsonl",
        split="train",
    )
    source_rows = len(ds)

    ds = ds.shuffle(seed=SHUFFLE_SEED)
    ds = ds.filter(has_nonempty_contents)
    passed_nonempty = len(ds)
    logger.info(
        "lang=%s %d rows have both old+new contents (from %d)",
        lang, passed_nonempty, source_rows,
    )

    kept_rows: list[dict] = []
    passed_sha = passed_msg = passed_hash = passed_ngram = 0

    for row in tqdm(ds, desc=f"filter[{lang}]"):
        if row.get("commit") in eval_shas:
            continue
        passed_sha += 1

        instruction = (row.get("message") or "").strip()
        if not instruction:
            continue
        passed_msg += 1

        old_contents = row.get("old_contents") or ""
        new_contents = row.get("new_contents") or ""
        diff = build_unified_diff(
            old_contents,
            new_contents,
            old_file=row.get("old_file", "a"),
            new_file=row.get("new_file", "b"),
        )

        reason = contamination_reason(
            instruction, diff, old_contents, new_contents,
            eval_field_hashes, eval_ngrams,
            n=ngram_n, threshold=overlap_threshold,
        )
        if reason == "hash":
            continue
        passed_hash += 1
        if reason == "ngram":
            continue
        passed_ngram += 1

        if max_per_lang and len(kept_rows) >= max_per_lang:
            continue
        kept_rows.append({
            "instruction": instruction,
            "diff": diff,
            "before": old_contents,
            "after": new_contents,
            "commit": row.get("commit"),
        })

    logger.info(
        "lang=%s passed_nonempty=%d passed_sha=%d passed_msg=%d "
        "passed_hash=%d passed_ngram=%d kept=%d",
        lang, passed_nonempty, passed_sha, passed_msg,
        passed_hash, passed_ngram, len(kept_rows),
    )

    config_name = lang.replace("+", "p").replace("#", "sharp")
    logger.info("lang=%s pushing to %s (config=%s)", lang, output_repo, config_name)
    Dataset.from_list(kept_rows).push_to_hub(
        repo_id=output_repo,
        config_name=config_name,
        private=True,
    )

    return {
        "config": config_name,
        "funnel": {
            "source_rows": source_rows,
            "passed_nonempty": passed_nonempty,
            "passed_sha": passed_sha,
            "passed_nonempty_msg": passed_msg,
            "passed_hash": passed_hash,
            "passed_ngram": passed_ngram,
            "kept": len(kept_rows),
        },
    }


def main(argv: list[str] | None = None) -> int:
    """CLI entrypoint: build the decontaminated training set for each language and write the run report."""
    parser = argparse.ArgumentParser(description="Build CodeEditSearch training set.")
    parser.add_argument("--output-repo", default="lightonai/CodeEditSearchTrain",
                        help="HF Hub dataset repo to push to")
    parser.add_argument("--languages", nargs="+", default=None,
                        help="languages to build (default: auto-discover from source via --min-rows)")
    parser.add_argument("--min-rows", type=int, default=1000,
                        help="when auto-discovering, only include configs with >= this many rows (default: 1000)")
    parser.add_argument("--max-per-lang", type=int, default=None,
                        help="cap rows per language (default: no cap)")
    parser.add_argument("--overlap-threshold", type=float, default=0.5,
                        help="13-gram containment threshold (default 0.5)")
    parser.add_argument("--ngram-n", type=int, default=13, help="n-gram size (default 13)")
    args = parser.parse_args(argv)

    logging.basicConfig(
        level=logging.INFO,
        format="%(asctime)s %(levelname)s %(name)s: %(message)s",
        stream=sys.stderr,
    )

    load_dotenv()
    assert os.environ.get("HF_TOKEN"), "HF_TOKEN not set after load_dotenv(); add it to .env"

    if args.languages is None:
        args.languages = discover_languages(SOURCE_DATASET, args.min_rows)
    logger.info("building %d languages: %s", len(args.languages), args.languages)

    # The eval (cassanof/CodeEditSearch) only has the 13 EVAL_LANGUAGES configs,
    # so only query those — intersected with what the user requested.
    eval_overlap_langs = [l for l in args.languages if l in EVAL_LANGUAGES]
    eval_shas = collect_eval_shas(eval_overlap_langs)
    eval_field_hashes, eval_ngrams = build_eval_overlap_index(
        eval_overlap_langs, n=args.ngram_n,
    )

    report = {
        "source_dataset": SOURCE_DATASET,
        "eval_dataset": EVAL_DATASET,
        "shuffle_seed": SHUFFLE_SEED,
        "overlap_threshold": args.overlap_threshold,
        "ngram_n": args.ngram_n,
        "eval_shas_count": len(eval_shas),
        "eval_field_hashes_count": len(eval_field_hashes),
        "eval_ngrams_count": len(eval_ngrams),
        "languages": [],
    }
    for lang in args.languages:
        report["languages"].append(build_language(
            lang=lang,
            eval_shas=eval_shas,
            eval_field_hashes=eval_field_hashes,
            eval_ngrams=eval_ngrams,
            output_repo=args.output_repo,
            max_per_lang=args.max_per_lang,
            ngram_n=args.ngram_n,
            overlap_threshold=args.overlap_threshold,
        ))

    totals: dict[str, int] = {}
    for entry in report["languages"]:
        for k, v in entry["funnel"].items():
            totals[k] = totals.get(k, 0) + v
    report["totals"] = totals

    REPORT_DIR.mkdir(parents=True, exist_ok=True)
    report_path = REPORT_DIR / "report.json"
    report_path.write_text(json.dumps(report, indent=2))
    logger.info("wrote report to %s", report_path)
    logger.info("totals: %s", json.dumps(totals))
    logger.info("pushed to https://huggingface.co/datasets/%s", args.output_repo)
    return 0


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

Filtering report

Per-language decontamination funnel

{
  "source_dataset": "bigcode/commitpackft",
  "eval_dataset": "cassanof/CodeEditSearch",
  "shuffle_seed": 42,
  "overlap_threshold": 0.5,
  "ngram_n": 13,
  "eval_shas_count": 21495,
  "eval_field_hashes_count": 84365,
  "eval_ngrams_count": 2887107,
  "languages": [
    {
      "config": "batchfile",
      "funnel": {
        "source_rows": 1466,
        "passed_nonempty": 1234,
        "passed_sha": 1234,
        "passed_nonempty_msg": 1234,
        "passed_hash": 1234,
        "passed_ngram": 1232,
        "kept": 1232
      }
    },
    {
      "config": "bitbake",
      "funnel": {
        "source_rows": 1308,
        "passed_nonempty": 924,
        "passed_sha": 924,
        "passed_nonempty_msg": 924,
        "passed_hash": 924,
        "passed_ngram": 924,
        "kept": 924
      }
    },
    {
      "config": "c",
      "funnel": {
        "source_rows": 8506,
        "passed_nonempty": 6925,
        "passed_sha": 5335,
        "passed_nonempty_msg": 5335,
        "passed_hash": 5027,
        "passed_ngram": 4415,
        "kept": 4415
      }
    },
    {
      "config": "csharp",
      "funnel": {
        "source_rows": 9346,
        "passed_nonempty": 8472,
        "passed_sha": 8472,
        "passed_nonempty_msg": 8472,
        "passed_hash": 8467,
        "passed_ngram": 8358,
        "kept": 8358
      }
    },
    {
      "config": "cpp",
      "funnel": {
        "source_rows": 4992,
        "passed_nonempty": 3855,
        "passed_sha": 2165,
        "passed_nonempty_msg": 2165,
        "passed_hash": 1967,
        "passed_ngram": 1748,
        "kept": 1748
      }
    },
    {
      "config": "clojure",
      "funnel": {
        "source_rows": 2403,
        "passed_nonempty": 2194,
        "passed_sha": 2194,
        "passed_nonempty_msg": 2194,
        "passed_hash": 2191,
        "passed_ngram": 2191,
        "kept": 2191
      }
    },
    {
      "config": "coffeescript",
      "funnel": {
        "source_rows": 5513,
        "passed_nonempty": 5131,
        "passed_sha": 5131,
        "passed_nonempty_msg": 5131,
        "passed_hash": 5129,
        "passed_ngram": 5116,
        "kept": 5116
      }
    },
    {
      "config": "css",
      "funnel": {
        "source_rows": 5049,
        "passed_nonempty": 4762,
        "passed_sha": 4762,
        "passed_nonempty_msg": 4762,
        "passed_hash": 4762,
        "passed_ngram": 4705,
        "kept": 4705
      }
    },
    {
      "config": "elixir",
      "funnel": {
        "source_rows": 1150,
        "passed_nonempty": 1016,
        "passed_sha": 1016,
        "passed_nonempty_msg": 1016,
        "passed_hash": 1015,
        "passed_ngram": 1015,
        "kept": 1015
      }
    },
    {
      "config": "emacs-lisp",
      "funnel": {
        "source_rows": 1015,
        "passed_nonempty": 950,
        "passed_sha": 950,
        "passed_nonempty_msg": 950,
        "passed_hash": 950,
        "passed_ngram": 950,
        "kept": 950
      }
    },
    {
      "config": "go",
      "funnel": {
        "source_rows": 5004,
        "passed_nonempty": 5001,
        "passed_sha": 3249,
        "passed_nonempty_msg": 3249,
        "passed_hash": 3097,
        "passed_ngram": 2925,
        "kept": 2925
      }
    },
    {
      "config": "groovy",
      "funnel": {
        "source_rows": 1486,
        "passed_nonempty": 1026,
        "passed_sha": 1026,
        "passed_nonempty_msg": 1026,
        "passed_hash": 1026,
        "passed_ngram": 832,
        "kept": 832
      }
    },
    {
      "config": "haml",
      "funnel": {
        "source_rows": 4415,
        "passed_nonempty": 4346,
        "passed_sha": 4346,
        "passed_nonempty_msg": 4346,
        "passed_hash": 4346,
        "passed_ngram": 4344,
        "kept": 4344
      }
    },
    {
      "config": "handlebars",
      "funnel": {
        "source_rows": 1429,
        "passed_nonempty": 1389,
        "passed_sha": 1389,
        "passed_nonempty_msg": 1389,
        "passed_hash": 1389,
        "passed_ngram": 1389,
        "kept": 1389
      }
    },
    {
      "config": "haskell",
      "funnel": {
        "source_rows": 1389,
        "passed_nonempty": 1389,
        "passed_sha": 1389,
        "passed_nonempty_msg": 1389,
        "passed_hash": 1388,
        "passed_ngram": 1385,
        "kept": 1385
      }
    },
    {
      "config": "html",
      "funnel": {
        "source_rows": 20214,
        "passed_nonempty": 18303,
        "passed_sha": 18303,
        "passed_nonempty_msg": 18303,
        "passed_hash": 18287,
        "passed_ngram": 18169,
        "kept": 18169
      }
    },
    {
      "config": "htmlperb",
      "funnel": {
        "source_rows": 10910,
        "passed_nonempty": 10458,
        "passed_sha": 10458,
        "passed_nonempty_msg": 10458,
        "passed_hash": 10456,
        "passed_ngram": 10438,
        "kept": 10438
      }
    },
    {
      "config": "ini",
      "funnel": {
        "source_rows": 11360,
        "passed_nonempty": 10626,
        "passed_sha": 10626,
        "passed_nonempty_msg": 10626,
        "passed_hash": 10553,
        "passed_ngram": 10295,
        "kept": 10295
      }
    },
    {
      "config": "jade",
      "funnel": {
        "source_rows": 1119,
        "passed_nonempty": 1098,
        "passed_sha": 1098,
        "passed_nonempty_msg": 1098,
        "passed_hash": 1097,
        "passed_ngram": 1096,
        "kept": 1096
      }
    },
    {
      "config": "java",
      "funnel": {
        "source_rows": 20635,
        "passed_nonempty": 13579,
        "passed_sha": 11823,
        "passed_nonempty_msg": 11823,
        "passed_hash": 11571,
        "passed_ngram": 9831,
        "kept": 9831
      }
    },
    {
      "config": "javascript",
      "funnel": {
        "source_rows": 52989,
        "passed_nonempty": 52235,
        "passed_sha": 50524,
        "passed_nonempty_msg": 50524,
        "passed_hash": 50124,
        "passed_ngram": 49310,
        "kept": 49310
      }
    },
    {
      "config": "json",
      "funnel": {
        "source_rows": 39777,
        "passed_nonempty": 38065,
        "passed_sha": 38065,
        "passed_nonempty_msg": 38065,
        "passed_hash": 38025,
        "passed_ngram": 38023,
        "kept": 38023
      }
    },
    {
      "config": "jsx",
      "funnel": {
        "source_rows": 2199,
        "passed_nonempty": 2056,
        "passed_sha": 2056,
        "passed_nonempty_msg": 2056,
        "passed_hash": 2056,
        "passed_ngram": 2048,
        "kept": 2048
      }
    },
    {
      "config": "kotlin",
      "funnel": {
        "source_rows": 2214,
        "passed_nonempty": 1838,
        "passed_sha": 1838,
        "passed_nonempty_msg": 1838,
        "passed_hash": 1837,
        "passed_ngram": 1607,
        "kept": 1607
      }
    },
    {
      "config": "less",
      "funnel": {
        "source_rows": 1360,
        "passed_nonempty": 1340,
        "passed_sha": 1340,
        "passed_nonempty_msg": 1340,
        "passed_hash": 1340,
        "passed_ngram": 1335,
        "kept": 1335
      }
    },
    {
      "config": "markdown",
      "funnel": {
        "source_rows": 62518,
        "passed_nonempty": 52850,
        "passed_sha": 52850,
        "passed_nonempty_msg": 52850,
        "passed_hash": 52825,
        "passed_ngram": 52258,
        "kept": 52258
      }
    },
    {
      "config": "nix",
      "funnel": {
        "source_rows": 1593,
        "passed_nonempty": 1551,
        "passed_sha": 1551,
        "passed_nonempty_msg": 1551,
        "passed_hash": 1551,
        "passed_ngram": 1551,
        "kept": 1551
      }
    },
    {
      "config": "perl",
      "funnel": {
        "source_rows": 2288,
        "passed_nonempty": 1453,
        "passed_sha": 1453,
        "passed_nonempty_msg": 1453,
        "passed_hash": 1451,
        "passed_ngram": 1435,
        "kept": 1435
      }
    },
    {
      "config": "php",
      "funnel": {
        "source_rows": 24791,
        "passed_nonempty": 20038,
        "passed_sha": 18293,
        "passed_nonempty_msg": 18293,
        "passed_hash": 17830,
        "passed_ngram": 17184,
        "kept": 17184
      }
    },
    {
      "config": "python",
      "funnel": {
        "source_rows": 56025,
        "passed_nonempty": 42537,
        "passed_sha": 40892,
        "passed_nonempty_msg": 40892,
        "passed_hash": 40393,
        "passed_ngram": 38700,
        "kept": 38700
      }
    },
    {
      "config": "restructuredtext",
      "funnel": {
        "source_rows": 6560,
        "passed_nonempty": 6011,
        "passed_sha": 6011,
        "passed_nonempty_msg": 6011,
        "passed_hash": 6010,
        "passed_ngram": 5978,
        "kept": 5978
      }
    },
    {
      "config": "ruby",
      "funnel": {
        "source_rows": 69413,
        "passed_nonempty": 62121,
        "passed_sha": 60504,
        "passed_nonempty_msg": 60504,
        "passed_hash": 59872,
        "passed_ngram": 58154,
        "kept": 58154
      }
    },
    {
      "config": "rust",
      "funnel": {
        "source_rows": 2996,
        "passed_nonempty": 2341,
        "passed_sha": 646,
        "passed_nonempty_msg": 646,
        "passed_hash": 598,
        "passed_ngram": 533,
        "kept": 533
      }
    },
    {
      "config": "scala",
      "funnel": {
        "source_rows": 5040,
        "passed_nonempty": 4255,
        "passed_sha": 2790,
        "passed_nonempty_msg": 2790,
        "passed_hash": 2287,
        "passed_ngram": 1963,
        "kept": 1963
      }
    },
    {
      "config": "scss",
      "funnel": {
        "source_rows": 6829,
        "passed_nonempty": 6796,
        "passed_sha": 6796,
        "passed_nonempty_msg": 6796,
        "passed_hash": 6796,
        "passed_ngram": 6733,
        "kept": 6733
      }
    },
    {
      "config": "shell",
      "funnel": {
        "source_rows": 31217,
        "passed_nonempty": 26026,
        "passed_sha": 24624,
        "passed_nonempty_msg": 24624,
        "passed_hash": 23969,
        "passed_ngram": 23040,
        "kept": 23040
      }
    },
    {
      "config": "slim",
      "funnel": {
        "source_rows": 1052,
        "passed_nonempty": 1034,
        "passed_sha": 1034,
        "passed_nonempty_msg": 1034,
        "passed_hash": 1033,
        "passed_ngram": 1032,
        "kept": 1032
      }
    },
    {
      "config": "sql",
      "funnel": {
        "source_rows": 2069,
        "passed_nonempty": 1211,
        "passed_sha": 1211,
        "passed_nonempty_msg": 1211,
        "passed_hash": 1206,
        "passed_ngram": 1187,
        "kept": 1187
      }
    },
    {
      "config": "swift",
      "funnel": {
        "source_rows": 4849,
        "passed_nonempty": 3955,
        "passed_sha": 2233,
        "passed_nonempty_msg": 2233,
        "passed_hash": 2079,
        "passed_ngram": 1870,
        "kept": 1870
      }
    },
    {
      "config": "text",
      "funnel": {
        "source_rows": 46588,
        "passed_nonempty": 44799,
        "passed_sha": 44799,
        "passed_nonempty_msg": 44799,
        "passed_hash": 44797,
        "passed_ngram": 44266,
        "kept": 44266
      }
    },
    {
      "config": "toml",
      "funnel": {
        "source_rows": 3424,
        "passed_nonempty": 3380,
        "passed_sha": 3380,
        "passed_nonempty_msg": 3380,
        "passed_hash": 3377,
        "passed_ngram": 3368,
        "kept": 3368
      }
    },
    {
      "config": "twig",
      "funnel": {
        "source_rows": 1610,
        "passed_nonempty": 1479,
        "passed_sha": 1479,
        "passed_nonempty_msg": 1479,
        "passed_hash": 1479,
        "passed_ngram": 1477,
        "kept": 1477
      }
    },
    {
      "config": "typescript",
      "funnel": {
        "source_rows": 5868,
        "passed_nonempty": 5328,
        "passed_sha": 3623,
        "passed_nonempty_msg": 3623,
        "passed_hash": 3488,
        "passed_ngram": 3371,
        "kept": 3371
      }
    },
    {
      "config": "unknown",
      "funnel": {
        "source_rows": 1597,
        "passed_nonempty": 1457,
        "passed_sha": 1457,
        "passed_nonempty_msg": 1457,
        "passed_hash": 1457,
        "passed_ngram": 1445,
        "kept": 1445
      }
    },
    {
      "config": "viml",
      "funnel": {
        "source_rows": 1063,
        "passed_nonempty": 1063,
        "passed_sha": 1063,
        "passed_nonempty_msg": 1063,
        "passed_hash": 1062,
        "passed_ngram": 1062,
        "kept": 1062
      }
    },
    {
      "config": "xml",
      "funnel": {
        "source_rows": 9337,
        "passed_nonempty": 8113,
        "passed_sha": 8113,
        "passed_nonempty_msg": 8113,
        "passed_hash": 8109,
        "passed_ngram": 7700,
        "kept": 7700
      }
    },
    {
      "config": "yaml",
      "funnel": {
        "source_rows": 114320,
        "passed_nonempty": 103688,
        "passed_sha": 103688,
        "passed_nonempty_msg": 103688,
        "passed_hash": 103666,
        "passed_ngram": 102934,
        "kept": 102934
      }
    }
  ],
  "totals": {
    "source_rows": 678295,
    "passed_nonempty": 599698,
    "passed_sha": 578203,
    "passed_nonempty_msg": 578203,
    "passed_hash": 573593,
    "passed_ngram": 560922,
    "kept": 560922
  }
}

Licensing

This dataset is derived from bigcode/commitpackft and we keep all of the original licenses of the source data.

Citation

If you use this dataset, please consider citing our work

@misc{sourty2026denseonlateonfullyopen,
  title         = {DenseOn with the LateOn: Fully Open Dense and Late-Interaction Models for Multilingual, Long-Context, and Code Search},
  author        = {Raphaël Sourty and Antoine Chaffin and Paulo Roberto Moura Junior and Amélie Chatelain},
  year          = {2026},
  eprint        = {2607.27178},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2607.27178},
}