CiberIA PromptRisk v1 — reproducibility and corpus provenance evidence

#1
by salim4n - opened

Hello Jordi,

We are reviewing CiberIA PromptRisk v1 for a fresh independent document-sanitization evaluation. We have not imported its rows or tested our candidate against them.

Our review targets this exact dataset revision:
https://huggingface.co/datasets/gcjordi/CiberIA-PromptRisk-v1/tree/c6ada043f9b1788927e53a2f0d4c957c205bd809

The card says the source package includes a deterministic generator, a validation report and a TF-IDF baseline. We could not locate that package in the dataset repository, the associated PromptGuard model repository, or the public GitHub repositories/search we reviewed.

Could you provide:

  1. A versioned source-package URL or archive, with a commit/release identifier, generator code, original template/pattern inventory, dependencies, seed/configuration and reproduction command corresponding to that exact dataset revision.
  2. The dataset-construction validation report, including file hashes and evidence that the generator reproduces the published Parquet shards and their pattern/scenario/translation split isolation. We found the model evaluation results; the missing evidence concerns dataset construction.
  3. An explicit dated authorship statement identifying whether any LLM generated, paraphrased, translated, augmented or labelled any base patterns, final texts or translations. Deterministic expansion alone does not establish human authorship of its inputs. If some material used an LLM, please identify its record/template IDs and whether a fully human-authored subset is traceable.
  4. The record-to-template/source lineage and any third-party text sources or licence exceptions. The public dataset licence is CC-BY-4.0; please confirm that it covers all selected texts and translations and give the source-code licence separately.
  5. Any existing malicious/benign counterpart relation. If no reciprocal pair mapping exists, please confirm that; we can manually document property-based pairing of independently authored rows without rewriting them.

Please link the evidence to revision c6ada043f9b1788927e53a2f0d4c957c205bd809. An inspectable public package is preferable; no proprietary audit battery, private scoring rules, credentials, customer data or new benchmark text is requested.

Thank you.

Owner

Thank you for the careful review. Your observation is correct: the public Hugging Face dataset repository currently does not contain the full construction/reproducibility package referred to in the Dataset Card.

The original PromptRisk v1 development package included the dataset-generation/validation workflow, leakage checks, a TF-IDF baseline, compatibility information and checksum validation, but those supporting artifacts were not all published alongside the dataset repository. We will correct that documentation/publication gap and centralize the reproducibility material in the dataset repository.

A few clarifications in advance:

  • PromptRisk v1 is a synthetic, human-directed dataset. AI assistance was used during the dataset-development process, so we will not characterize the source material as exclusively human-authored. We will add an explicit authorship/AI-assistance statement describing the construction process and the limits of the provenance information available for v1.

  • The published records were produced through a controlled deterministic construction process and grouped using pattern_group / scenario_group so that translations, scenarios and underlying template patterns do not cross train/validation/test splits.

  • We will publish the reproducibility material available from the original v1 package, including generation/validation assets, configuration/dependency information, checksums and the TF-IDF baseline, and clearly distinguish original release artifacts from any documentation added retrospectively.

  • The dataset license remains CC BY 4.0; code/reproducibility tooling will be separately licensed under Apache-2.0 as stated in the Dataset Card.

  • PromptRisk v1 does not define an explicit one-to-one malicious ↔ benign counterpart relation between records. The benign side includes deliberately constructed hard negatives, but these should not be interpreted as formal paired counterfactuals unless explicitly identified as such.

We appreciate the reproducibility review, especially since the dataset is being considered for an independent evaluation. We will update this thread with the exact revision and reproducibility-package location once the repository documentation has been aligned with the published claims.

Owner

Following up on my previous comment: the reproducibility addendum is now publicly available on the current main revision, d25fd20.

It adds:

  • validation code, a generated validation report and SHA-256 manifests for the published Parquet shards;

  • checks for schema, counts, class balance, ID/text uniqueness, multilingual scenario completeness, and isolation of scenario_group, pattern_group and template_id across train, validation and test;

  • a transparent deterministic TF-IDF + logistic-regression baseline with machine-readable results;

  • explicit authorship/AI-use, provenance, release-status and code-licensing statements.

The published corpus passes all 16 validation checks, and the Parquet files themselves are unchanged.

To clarify the remaining provenance points: PromptRisk v1 is a synthetic, human-directed, AI-assisted research dataset; it is not claimed to be wholly human-authored, and no 100% human-only subset is asserted. The addendum validates the published corpus but does not claim retrospective byte-for-byte regeneration of the unpublished original authoring/generation package. The original templates, generation configuration and complete record-level audit trail were not part of the public v1 release; this limitation is now stated explicitly.

All rows are marked as synthetic, with no third-party text corpus, customer data, real credentials or production data. The dataset remains CC BY 4.0 and the reproducibility code is separately licensed under Apache-2.0. Finally, v1 does not define formal one-to-one malicious ↔ benign counterpart pairs; its benign hard negatives are related examples, not counterfactual pairings.

Thank you again for the review. I hope this makes the released corpus suitable for your independent evaluation.

Owner

Closing this discussion as the requested reproducibility, validation and provenance documentation has now been published and the relevant limitations of the v1 release have been explicitly documented.

Thank you again for the careful review. If further clarification is needed after reviewing the addendum, please feel free to open a new discussion.

gcjordi changed discussion status to closed

Sign up or log in to comment