| --- |
| license: apache-2.0 |
| pretty_name: KRAZYKITTY // DATA LAB |
| tags: |
| - data-provenance |
| - dataset-documentation |
| - preprocessing |
| - data-quality |
| - evaluation |
| - privacy |
| - reproducibility |
| --- |
| |
| # KRAZYKITTY // DATA LAB |
|
|
| **Status: research documentation and workflow toolkit — no dataset is published here.** |
|
|
| DATA LAB is KrazyKitty's evidence-first workspace for studying how datasets should be sourced, |
| reviewed, transformed, documented, and evaluated before they are allowed near a training or |
| benchmark pipeline. |
|
|
| This repository contains dataset-card templates, provenance schemas, deterministic preprocessing |
| tools, evaluation guidance, and review gates. It does **not** contain a corpus, benchmark examples, |
| training rows, scraped content, or a claim that any source is approved for KittyLM. |
|
|
| ## Research tracks |
|
|
| | Track | Purpose | Current state | |
| |---|---|---| |
| | Source provenance | Bind each source to identity, version, integrity, licence, and allowed use | Working templates and validator | |
| | Licensing, privacy, consent | Require explicit permission and data-risk review before processing | Documented gate | |
| | Preprocessing experiments | Compare deterministic text transformations without hiding exclusions | Working local reference workflow | |
| | Data quality | Measure duplication, encoding, language, structure, contamination, and coverage | Measurement plan | |
| | Evaluation resources | Version benchmark inputs and separate test integrity from model claims | Documentation and Collection candidates | |
| | Dataset cards | Make intended use, composition, limitations, and governance visible | Reusable templates | |
|
|
| ## Repository map |
|
|
| - [`docs/PROVENANCE.md`](docs/PROVENANCE.md) — evidence model and source lineage |
| - [`docs/LICENSE_PRIVACY_CONSENT.md`](docs/LICENSE_PRIVACY_CONSENT.md) — release-blocking review |
| - [`docs/DATA_QUALITY.md`](docs/DATA_QUALITY.md) — quality dimensions and reporting rules |
| - [`docs/EVALUATION.md`](docs/EVALUATION.md) — benchmark governance and contamination controls |
| - [`templates/dataset-card-template.md`](templates/dataset-card-template.md) — corpus-oriented card |
| - [`templates/evaluation-card-template.md`](templates/evaluation-card-template.md) — evaluation-set card |
| - [`schemas/source-manifest.schema.json`](schemas/source-manifest.schema.json) — machine-readable contract |
| - [`workflows/validate_manifest.py`](workflows/validate_manifest.py) — standard-library validator |
| - [`workflows/preprocess_text.py`](workflows/preprocess_text.py) — deterministic, approval-gated text reference flow |
| - [`experiments/README.md`](experiments/README.md) — experiment record contract |
|
|
| ## Reproducible workflow |
|
|
| 1. Create a source manifest from the schema and keep its decision `proposed`. |
| 2. Record canonical source URLs, versions, licences, redistribution terms, integrity hashes, |
| privacy review, consent or other lawful basis, and intended use. |
| 3. Review the source outside the preprocessing program. Automated checks do not grant permission. |
| 4. Change the manifest to `approved` only after the named reviewers have approved every source. |
| 5. Run `python workflows/validate_manifest.py manifest.json`. |
| 6. Run the reference preprocessing experiment against only the files listed in that manifest. |
| 7. Store the exact command, source revision, manifest hash, environment, counts, exclusions, and |
| output hashes in an experiment record. |
| 8. Publish data only through a separate release decision with a complete dataset card. |
|
|
| ## Non-goals and safeguards |
|
|
| - Public availability is not treated as consent or permission for model training. |
| - A licence label is not accepted without checking scope, attribution, redistribution, and |
| downstream restrictions. |
| - The reference preprocessing tool is not a privacy scrubber, copyright detector, toxicity |
| classifier, or quality oracle. |
| - Personally identifiable, sensitive, confidential, or revoked material is release-blocking until |
| reviewed and resolved. |
| - Benchmark test content must not enter training mixtures. |
| - Raw and processed data remain outside this repository unless a later, separately approved |
| dataset release explicitly permits them. |
|
|
| ## Relationship to KittyLM |
|
|
| DATA LAB can produce reviewed methods and provenance records that KittyLM may use in the future. |
| It does not imply that KittyLM has an approved training dataset, completed data pipeline, or |
| published model weights. |
|
|
| ## Licence |
|
|
| The documentation, schemas, and original workflow code in this repository are proposed for release |
| under Apache-2.0. That licence does not apply to any third-party source described by a manifest. |
| Every source retains its own rights and restrictions. |
|
|
