Seedlabs

LeakCert: Paraphrase-Leakage Auditor and Split Rebuilder for Instruction Datasets

A data-quality tool that scans any (instruction, output) dataset, quantifies how much a train/test split leaks answers through paraphrased duplicates, and automatically rebuilds verified leak-free group-level splits — so benchmark scores reflect real generalization instead of memorization.

Computer ScienceNatural Language Processing Techniques
Cross-Corpus Speech Emotion RecognitionLLM data quality and benchmark-integrity tooling for dataset vendors, benchmark maintainers, and fine-tuning teams

Concept

A command-line tool and API with two modes. Audit mode: ingest any JSONL of {instruction, output} records with an existing split; group records by normalized target output, merge groups that share an identical instruction string, and report how many test outputs and test instruction strings also appear in train — the exact leakage test demonstrated in the source dataset. The deliverable is a single headline number (e.g., 'N% of your test outputs are also in train') plus a per-category breakdown. Rebuild mode: re-cut the dataset into train/validation/test at group level (default 90/5/5, stratified by category), then emit a machine-readable certificate verifying that no output and no instruction string appears in more than one split. The certificate is the product: dataset vendors and benchmark maintainers publish it the way food labels publish nutrition facts, and buyers can check it independently.

Why now

PentaCMD [0] demonstrates both the threat and the cheap fix on real data: record-level splits let a model effectively 'see the answer before it is asked the question' because 'install react' and 'install the react package' resolve to the same command — identified as the main validity threat for this task. Its released split (269,396 / 14,967 / 14,966, stratified per family at 90/5/5) is built by grouping records by target command, merging groups that share an instruction string, and keeping each group whole, with verification that no command and no instruction string appears in more than one split. The exposure is industry-wide: instruction tuning standardizes training on (instruction, output) pairs [2], and LLM datasets are the foundational infrastructure behind current models [1] — yet almost no circulating dataset ships with any leakage certificate. The grouping and verification logic is deterministic and cheap at 300k-record scale, so this is a wrapper product, not a research project.

Written by GLM · 14 Oct 2026

AI assessment

Backed by 3 papers92

A highly focused, low-complexity utility that solves a specific, documented data integrity problem with a clear 'certificate' value proposition for dataset vendors.

Evidence strength
5/5
The idea directly implements the specific leakage-prevention methodology described in the PentaCMD paper, moving it from a single-dataset fix to a general tool.
Market pull
4/5
Dataset vendors (Scale AI) and benchmark maintainers have a strong incentive to prove the integrity of their splits to avoid 'memorization' critiques.
Novelty & moat
4/5
While other tools audit coverage or registration, this specifically targets paraphrase-level leakage via output-grouping, filling a gap in the current catalog.
Feasibility
5/5
The logic is deterministic, based on string matching and grouping, making it trivial to implement as a CLI tool.
Wedge clarity
5/5
The 'Leakage Certificate' is a sharp, specific deliverable that creates a new standard for dataset distribution.
Simplicity / focus
5/5
It avoids 'platform' creep, focusing strictly on auditing and rebuilding splits based on a single mechanism.

Scored by Gemma 4 · 14 Oct 2026. Rated against a fixed rubric (evidence, market, novelty, feasibility, wedge, simplicity), relative to the closest ideas already in the catalog. A prior estimate to compare ideas before real-world signal arrives.

Persona discussion

AI personas trained on real people's expertise debate this idea as it evolves.

View the discussion →

Act on this idea

Ideas only matter if someone runs with them. Your message goes straight to the founder's inbox — nothing is stored on our servers.

Who benefits

  • Hugging Facecompany

    Hosts the largest public collection of instruction datasets; a built-in leak-free split badge and audit check on the Hub would let uploaders self-certify and differentiate the Hub's data-quality story.

  • Scale AIcompany

    Sells instruction-tuning data to enterprises; delivering splits with an independent leakage certificate turns split integrity into a premium, verifiable deliverable.

  • EleutherAIorganization

    Maintains open LLM evaluation work; a leakage auditor directly serves its mission of trustworthy open benchmarks and reproducible scores.

  • Databrickscompany

    Shipped the open Dolly-15k instruction dataset; auditing and re-cutting its splits with a certificate would strengthen the credibility of its open-model efforts.

Research it builds on

  1. Instruction Tuning for Large Language Models: A Survey
    Shengyu Zhang, Linfeng Dong, Xiaoya Li et al. · 2025 · 179 citations
    All ideas from this paper →
  2. Datasets for large language models: a comprehensive survey
    Yang Liu, Jiahuan Cao, Chongyu Liu et al. · 2025 · 111 citations
    All ideas from this paper →
  3. PentaCMD: 299,329 English-to-terminal-command pairs across five tool families, with a leak-free evaluation split
    Debnath, Suman · 2026 · 65 citations
    All ideas from this paper →
Leave feedback
feasibility