Home/Docs/Training-data flywheel
Follow-up · evidence to training

Turn measured harness lift into reviewable training data.

The evaluation harness can do more than score a model. Comparable stock and harnessed answers, citations, tool traces, and per-dimension grades can become candidate teaching material. DueCare turns that evidence into lineage-safe supervised fine-tuning (SFT) data first, Direct Preference Optimization (DPO) data second, and model-specific Low-Rank Adaptation (LoRA) experiments — without treating generated text as automatically fit to train.

Public release boundary · verified 2026-07-15. Two advanced Kaggle datasets are public with exact approval, privacy, license, checksum, and split-isolation evidence. The measured-response release contains 791 supervised fine-tuning rows, 791 preference pairs, 1,582 reward labels, and raw-text-free audit lanes. The multiperspective release contains 25,600 supervised fine-tuning training rows, 25,600 preference-training rows, and 2,048 rows in each held-out split. No Gemma fine-tuning, graphics-processing-unit run, production adapter, merged weights, or independent model-lift result is claimed.
Remote notebook evidence. Nine public Kaggle notebooks load, verify, explain, visualize, and run bounded central-processing-unit diagnostics on the releases. Downloaded Kaggle outputs contain 11 charts and 10 review tables for the response explorer, 15 charts and 7 review tables for the multiperspective explorer, plus loading, quality-dashboard, and small-classifier artifacts. The measured-response labels overlap the source benchmark, so its classifier results are diagnostics rather than independent improvement evidence.
00 · Start here

Public data that can be opened, loaded, and checked.

The landing route starts with the two versioned datasets, then moves through a loading guide and a cross-dataset quality audit before any training plan.

Measured response corpus

Positive supervised targets, same-prompt preference pairs, reward labels, a raw-text-free response inventory, and an explicit quarantine lane. Release-manifest Secure Hash Algorithm 256-bit checksum (SHA-256): 56fa69c…22b74b.

Multiperspective synthetic corpus

Deterministic fictional scenarios across perspective, journey stage, evidence state, time, and jurisdiction. It contains visible decision scaffolds, not real worker cases or hidden reasoning. Release-manifest SHA-256: ea644df4…9211d.

Loading quickstart

Verifies both manifests, explains each file type in plain language, demonstrates standard Python, pandas, Hugging Face Datasets, Kagglehub, and Polars loaders, and saves three reviewer-facing charts.

Quality dashboard

Checks exact prompt and lineage-family isolation across train, validation, and test splits; profiles text and shard sizes; and writes an audit table, summary, report, and four charts.

01 · The full flywheel

Every useful output stays evidence before it becomes data.

The same prompt is run through controlled arms, graded against the same rubric, and retained with exact identity and provenance. Promotion is a sequence of explicit gates, not a score threshold alone.

01 · RunComparable answers

Record stock and harnessed responses for an approved prompt, exact model revision, harness version, and knowledge-pack versions.

02 · JudgeDimension evidence

Retain citations, tool traces, deterministic checks, per-dimension grades, and intentionally visible reviewer rationales.

03 · GateEligible candidates

Reject unclear licensing, unsafe advice, sensitive data, unsupported citations, hidden-thought markup, duplicates, or incomplete grading.

04 · SplitFreeze lineages

Assign complete prompt and source lineages to train, validation, or held-out sets before trainer-specific formatting.

05 · SFTTeach the answer

Train on approved final answers and reviewable response structure. Stable behavior belongs in the adapter; volatile facts stay in tools.

06 · PreferenceTeach the boundary

After SFT, prefer a grounded chosen answer over a traceable rejected failure using DPO or another compatible preference method.

07 · EvaluateFour-arm proof

Compare stock, stock + harness, trained, and trained + harness on the untouched holdout and on benign controls.

08 · RepeatFeed back failures

Send reviewed gaps back to prompt design and curation. Do not train on the evaluation holdout that exposed them.

02 · Supervised fine-tuning, then preference

Teach an approved response before optimizing its preference.

Supervised fine-tuning and preference optimization answer different questions. DueCare records both stages separately so a skipped or failed preference stage cannot be reported as a completed training run.

Supervised fine-tuning

An SFT row maps an approved instruction or chat history to a reviewed final answer. It can include citations and a deliberately authored, model-visible reasoning scaffold. The prompt portion can be masked so the loss trains the reply rather than memorizing the input.

Preference optimization

A Direct Preference Optimization (DPO) row keeps one prompt with a chosen and rejected answer plus the reason for that preference. The rejected answer must represent a known failure — such as unsupported advice, missed evidence, or over-refusal — rather than an arbitrary weak sample.

Counterfactuals and hard negatives

Minimal-pair facts, controlled contract ablations, and reviewed reasoning repairs test whether the model follows evidence instead of keywords. These variants keep their source lineage and never silently replace the canonical row.

Chain-of-thought boundary

The export may contain complete final answers, citations, harness traces, concise explanations intentionally returned to the user, and visible judge or reviewer rationales. It does not retrieve, scrape, infer, or publish a provider’s private hidden chain-of-thought. Hidden-thought markup blocks a row from training.

03 · Dataset contract

Portable rows, immutable evidence.

Trainer formats can change. The canonical record stays tied to the prompt, source, model revision, review decision, and split assignment that made it eligible.

contract fieldwhat it recordswhy it blocks release when absent
IdentityStable row identifier, prompt hash, response hash, generator version, artifact hash.Prevents silent mutation and makes an exported row reproducible.
Lineage + splitPrompt lineage, source lineage, typology/corridor groups, train/validation/held-out assignment.Prevents the same family or a near duplicate from leaking into evaluation.
Model provenanceProvider, exact base model identifier, immutable revision, harness/rubric version, pack versions.An adapter and its result are meaningful only against the base revision actually used.
Source + licenseSource references, allowed use, license metadata, curator decision, transformation history.Public availability alone does not grant permission to train or redistribute.
Training contentFinal answer, citations, optional visible rationale, SFT messages or chosen/rejected preference pair.Separates reviewable model-visible material from prohibited hidden reasoning.
Safety evidencePersonally identifiable information scan, unsafe-advice check, citation verification, deterministic checks, grades, rejection reasons.A high aggregate score cannot override a privacy, evidence, or safety failure.
04 · Use cases

One evidence contract, several accountable workflows.

The flywheel is useful wherever a team needs to improve behavior without losing the ability to explain which data, model, and evaluation produced the change.

Harness maintainers

Turn recurring, reviewed failures into SFT targets, preference pairs, hard negatives, or counterfactual tests while preserving the original evaluation evidence.

Dataset curators

Review provenance, redistribution rights, citations, privacy findings, and lineage assignments before creating a versioned public training-data candidate.

Local Ollama red-team loop

Run adversarial rewrite, protective answer generation, and strict judge review against approved seed prompts on a local Ollama host. The output is supervised and preference candidate JavaScript Object Notation Lines plus raw-text-free quarantine metadata; it remains safe_to_train=false until the full release gates pass.

Model researchers

Compare how much harness behavior different compatible base models internalize, then measure whether the harness still adds value after training.

Independent reviewers

Recompute hashes, inspect exclusions, verify untouched held-out lineages, and distinguish a dataset release from a notebook run or adapter release.

05 · Kaggle and other systems

A-00 is the executable handoff, not a shortcut around review.

The active notebook can import an approved response bundle, export manifest-bound supervised and preference JavaScript Object Notation Lines, preflight a compatible NVIDIA Compute Unified Device Architecture runtime, run resumable Low-Rank Adaptation stages, and reload the exact base plus adapter for evaluation.

Gemma 4 on Kaggle

Start with a commit-pinned official Gemma 4 E2B or E4B model-variant preset, verify the dataset manifest and dependencies, run supervised fine-tuning, then run the requested preference stage. Download the adapter, logs, completion manifest, and four-arm results before the session ends. The larger 31-billion-parameter evaluation target is a workstation/server-class path, not a phone deployment claim.

Compatible model families

The JavaScript Object Notation Lines and manifest contracts can feed controlled Unsloth, Hugging Face Transformer Reinforcement Learning, or Parameter-Efficient Fine-Tuning environments for another compatible base. Each model and immutable revision receives its own adapter and evaluation; an adapter trained for one base is never presented as portable to another.

External importer

A-00’s Already have a file? path can inspect controlled JavaScript Object Notation, JavaScript Object Notation Lines, or compressed archive exports and suggest a prompt, grading, comparison, or training workflow. A loose file stays inspection-only: it cannot train until its supervised/preference artifacts and manifest pass the same hashes, model-revision, provenance, license, privacy, lineage, holdout, and quality gates.

Current Kaggle Dataset publication unit: eligible supervised fine-tuning, preference, reward-label, and declared audit JavaScript Object Notation Lines files; immutable manifest and checksums; Croissant metadata; loading guide; data card; source/license summary; split policy; gate report; and limitations. Raw cases, credentials, machine-local paths, private hidden reasoning, and training-excluded evaluation material do not belong in the release. Both public advanced datasets meet this packaging contract for their stated uses.
06 · Release ladder

Publish only the layer that has earned it.

A notebook, dataset version, and adapter are separate releases. Each carries its own evidence, and later layers cannot borrow confidence from earlier smoke tests.

  1. Complete exact generation and grading coverage. Partial sweeps remain resumable evidence, not a complete training corpus.
  2. Audit and curate candidate rows. Resolve provenance, redistribution rights, privacy, citations, safety, and incomplete judge coverage.
  3. Freeze a manifest-bound dataset version. Publish only rows eligible for the declared use, with hashes, lineage-safe splits, data card, and limitations. Both advanced public datasets passed this rung on 2026-07-15.
  4. Publish and verify a pinned notebook version. Bind input dataset version, code commit, model revision, dependencies, and output manifest; confirm the public Kaggle web address resolves before linking it.
  5. Train and evaluate. Complete SFT and the requested preference stage, then run the four arms on untouched held-out lineages.
  6. Consider an adapter release. Publish weights only if the intended dimensions improve without a citation, privacy, unsafe-assistance, or benign over-refusal regression.
Today: the repository implementation, two advanced public Kaggle datasets, nine public learning notebooks, and downloaded remote outputs are available. The dataset-publication rung passed. Gemma training, a production adapter, merged weights, and independently demonstrated model lift remain future rungs.
07 · Method and source

Review the contracts before running a graphics processing unit.

The website is an orientation layer. The repository owns the executable gates and the versioned methodology.