Formats
- Datasheet as Markdown and PDF
- Croissant as JSON-LD
- Register and manifest as CSV and Parquet, with content hashes
- Disclosure draft as Word or Markdown
Know where every training record came from, and be able to show it.
Regulators, customers, and your own counsel now ask the same question: what was this model trained on? We build the documentation that answers it from the dataset itself rather than from memory: a datasheet, machine-readable metadata, a source and license register, and a lineage manifest that traces each record from source system to training file.
The question used to come from researchers. Now it comes from statute.
California's AB 2013 requires developers of generative AI systems offered to Californians to post documentation about the data used to train them, with the first postings due by January 1, 2026. The required summary covers 12 enumerated items, among them the sources or owners of the datasets, the number of data points, whether the data includes copyrighted material or personal information, and whether synthetic data generation was used. The statute's definitions reach further than many enterprises expect: it treats retraining and fine-tuning as a substantial modification, and a developer includes anyone who substantially modifies a system for public use.
In the EU, Article 10 of the AI Act requires the training, validation, and testing data of high-risk systems to be governed by documented practices, and lists eight of them, from data collection and origin through annotation and cleaning to bias examination and gap analysis. Customer security questionnaires have started asking versions of the same questions.
None of this is hard to answer if the records were kept while the data was being prepared. It is slow and expensive to reconstruct a year later. Benner House is not a law firm and does not provide legal advice: we produce the evidence, and your counsel decides what it means and what to disclose.
The provenance package at a glance
You provide
- One dataset, or the set of sources a model was trained on
- Access to the pipelines, scripts, or notebooks that produced it
- Contracts, licenses, terms of use, and data-sharing agreements for each source
- A contact for each source system or data owner
We deliver
- A datasheet: motivation, composition, collection, preprocessing, uses, distribution, and maintenance
- Croissant metadata with the Responsible AI fields completed
- A source and license register stating the rights basis for each source
- A lineage manifest from source records to training files, with transformation and reviewer records
- A draft public disclosure organized by the AB 2013 items, for your counsel's review
Quality gates
- Every source traced to a contract, a license, or a documented internal origin, or flagged as unresolved
- Record counts in the manifest reconciled to the delivered files
- Content hashes recomputed and matched at delivery
- Disclosure draft checked line by line against the dataset by a second analyst
Turnaround
Two to three weeks per dataset
Team
A data engineer and a provenance analyst, working alongside your counsel
Starts at
$5,400per dataset
A datasheet, Croissant metadata, a source and license register, a lineage manifest, and a draft public disclosure for one dataset.
One set of evidence, mapped to each framework
We build the evidence once and map it to every framework that applies to you. The mapping is descriptive; what each obligation requires of you is your counsel's call.
| Framework | What it asks about data | Where the answer lives |
|---|---|---|
| California AB 2013 | A public, high-level summary of training datasets: sources, purpose, size, data types, IP and personal information, licensing, cleaning, collection period, synthetic data. | Disclosure draft, generated from the register and manifest |
| EU AI Act, Article 10 | Governance of training, validation, and testing data for high-risk systems: design choices, origin, preparation, assumptions, suitability, bias examination and mitigation, gaps. | Datasheet sections and the manifest's transformation log |
| NIST AI RMF and its Generative AI Profile | Documented data sources, provenance, and known limitations as part of mapping and measuring AI risk. | Register, datasheet, and evaluation results |
| Customer due diligence | Questionnaires on data rights, personal data, and whether customer data trains models. | Register and datasheet, answered once and reused |
| Internal model governance | Change control and audit trail from data version to model version. | Versioned manifest tied to each model release |
What a lineage manifest entry records
An illustrative entry, with invented values, for one source in a claims fine-tuning dataset. The full manifest has one of these per source and a hash for every delivered file. See a complete sample datasheet.
Source
- source_id
- SRC-014
- system
- Claims document management export, 2019–2025
- owner
- Claims operations
Why it matters: Someone can answer questions about this source
- rights_basis
- First-party business records; training use approved under internal data policy
Why it matters: Counsel reviewed the basis before training use
- records_in
- 412,880 documents
Why it matters: Reconciled to the export log, so nothing was lost or added
Transformations
- T1 parse
- Layout-aware OCR; tables kept as structured data
Why it matters: A pinned parser version makes the output reproducible
- T2 dedupe
- MinHash near-duplicate removal, threshold set per source
Why it matters: Every removed record is listed with a reason code
- T3 protect
- Identifiers replaced with consistent surrogates
Why it matters: The recall report is attached as evidence
- T4 label
- Coverage decisions reviewed against a versioned rubric
Why it matters: Reviewer IDs and agreement scores are attached
Delivery
- dataset_version
- claims-sft v1.3
Why it matters: A diff against v1.2 shows exactly what changed
- records_out
- 38,410 training examples
- integrity
- SHA-256 per file
Why it matters: Recomputed at delivery to show no file was altered
- synthetic
- None
Why it matters: Answers a disclosure question directly
Questions the package lets you answer in an afternoon
- Which sources went into model version 1.3, and which were dropped after 1.2?
- Does any training record come from a vendor file whose license excludes model training?
- Was personal information present, and what was done to it, by which rule version?
- How many records came from each source, and over what period were they collected?
- Who labeled the examples, under which rubric, and how often did reviewers agree?
- Did synthetic data enter the training set, and from which generator?
Priced per dataset
Provenance package
Starts at
$5,400per dataset
A datasheet, Croissant metadata, a source and license register, a lineage manifest, and a draft public disclosure for one dataset.
Request a scoped quoteEvery dataset Benner House delivers already ships with a lineage manifest. This package is for datasets prepared elsewhere or before you worked with us, or when you need the full disclosure set. Models trained on many datasets are quoted as a portfolio after a source inventory.
Questions about provenance
Are we a "developer" under AB 2013?
That is a legal question for your counsel. What we can describe is the text: the definition covers anyone who designs, codes, produces, or substantially modifies a generative AI system for public use, and it counts retraining and fine-tuning as substantial modification. We assemble the facts your counsel needs to decide.
Can you reconstruct provenance for a model that's already trained?
Partly. We rebuild lineage from pipeline code, job logs, storage snapshots, and interviews with the people who ran them. Where the evidence is gone, we mark the gap explicitly rather than filling it with assumptions, and we recommend what to log from now on.
What is Croissant, and why include it?
Croissant is an open metadata vocabulary for machine-learning datasets from MLCommons, built on schema.org. Hugging Face, Kaggle, and OpenML publish it, and its Responsible AI vocabulary records provenance and collection details in a form tools can read. It makes your documentation portable rather than locked in a PDF.
Will a public disclosure expose our trade secrets?
AB 2013 asks for a high-level summary, and it allows the number of data points to be stated in general ranges. We draft at the level of detail your counsel approves; the full register and manifest stay internal.
Does the manifest contain the data itself?
No. It holds identifiers, counts, hashes, rules, and reviewer records. The data stays in your environment, and the manifest can be shared with auditors without sharing any records.
Keep reading
Bring us one messy data source and one question.
Tell us what the data is and what you want a model to do with it. A data engineer, not a salesperson, will reply within one business day with questions and a proposed first step.