v0.4.0: Final testing, DSH confinement and reliable recovery

v0.4.0

New experiments reserve train/validation/final-test partitions. Training and Optuna see validation scores only; finalize commits one selected model per final partition before exposing final-test scores. Export and standalone retraining verify both metric sets. Existing two-way snapshots retain their original partitions and frozen recipes.

DSH now has 20 tools, including combined run, bounded wait and finalize. PTC retains complete canonical objects while Native rendering omits repetitive histories and detailed classification tables. Real DSH PTC, OS confinement and process-range shutdown were exercised; the implementation uses DSH services, not a replacement Agent executor.

Every new fit has a supervisor with a persisted wall-clock deadline. Trial timeouts work without polling. Local execution remains durable across DSH shutdown; the optional DSH execution mode applies current DSH file policy and stops owned workers on plugin/DSH shutdown. Missing confinement services fail closed. Windows reports partial file confinement, with no network/read-isolation claim.

An installation helper creates/reuses an isolated environment, installs locked dependencies, checks backend/plugin version equality, diagnoses dependencies/SQLite/Node, and generates a configuration snippet without overwriting the user's DSH profile.

Limits: local single-user CSV workflows; no hostile-code boundary, remote SSH backend, unattended search scheduler, nested CV, checkpoint resumption or general productivity claim. The earlier live Agent pilot remains a one-pair exploratory result. See validation evidence and maturity review.

DSH ML Workbench v0.3.0

DSH ML Workbench 0.3.0

Experiments now follow DSH Session and tool-call identity. ml_workbench_resume recovers run/study references; exact create/branch replays deduplicate automatically. Canonical result schemas expose typed PTC handles. Timeout messages explain recovery when a response is lost.

A live DSH pilot exposed rejection of redundant sklearn random_state and n_jobs values. Matching values are now normalized; conflicting controls remain errors. Both original failed payloads pass without duplicating runs.

Validation includes 31 distinct Python cases, four Node cases, Native-executor checks, four direct-MLflow comparisons and one live two-turn trial per arm. The pilot showed identical accuracies and successful continuation in both arms, but more tool calls with the plugin. The plugin delivery retrained its estimator; the direct workflow reloaded and rescored. The report explains the prompt's ambiguity and does not present this as an established productivity win.

Install both the 0.3.0 Python wheel and DSH bundle. Existing runs remain usable; old unlinked runs are accessible through ml_experiment_list. This is a local trusted-user adapter, not a full upstream-platform replacement or sandboxed training service.

DSH ML Workbench v0.2.0

0.2.0 — Durable Optuna tuning and stronger experiment workflows

DSH ML Workbench binds each local CSV experiment to its data, code, parameters, validation rows, MLflow record and fitted pipeline. A follow-up can reuse a prior experiment's exact fitted preprocessing while changing its estimator, without overwriting the original run. Incompatible evaluations and mismatched deliveries are rejected.

Sixteen DSH tools add dataset inspection, experiment diffs, batch prediction and durable Optuna search to the original experiment lifecycle. Group-disjoint and chronological holdouts, probability metrics, dummy baselines and confusion matrices improve evaluation. Deliveries now include an escaped portable HTML report.

MLflow 3.18.0, Optuna 5.0.0 and scikit-learn provide the underlying tracking, sampling, training and serialization. Optuna studies persist in SQLite with a linked MLflow run per trial. Repeated step calls do not duplicate active fits; lost controller responses are reconciled through persisted request keys. DVC and RD-Agent remain references rather than runtime integrations.

Validation covers the expanded Python regression/failure-injection suite, three Node tests, the actual 16-tool DSH executor, grouped/time synthetic fits, three Optuna trials, batch prediction, legacy v0.1 upgrade and independent delivery reproduction. See v0.2 validation evidence and its machine-readable results.

An online DSH account-provider call still returns ACCOUNT_SIGN_IN_REQUIRED before inference. Agent-productivity gains remain unmeasured. This remains a local single-user CSV workflow: Optuna requires step calls to advance and enforces timeout on polling; parallel tuning, pruning, nested CV, GPU training, remote serving and optimizer checkpoint continuation are not implemented. It is not feature-equivalent to the full upstream platforms.

Install the Python wheel with the supplied requirements.lock, then the DSH .tgz, or add this third-party index to DSH Discover:

https://raw.githubusercontent.com/shinerain1/dsh-ml-workbench/main/market/index.json

Configure absolute pythonExecutable and root paths as shown in the README. Only enable custom factories for reviewed local Python. This is an independent community plugin/index, not an official approval claim.

DSH ML Workbench v0.1.0

0.1.0 — MLflow workflows adapted for DSH

DSH ML Workbench binds each local CSV experiment to its data, code, parameters, validation rows, MLflow record and fitted pipeline. A follow-up can reuse a prior experiment's exact fitted preprocessing while changing its estimator, without overwriting the original run. Incompatible evaluations and mismatched deliveries are rejected.

Nine DSH tools cover health, create, start, get, list, cancel, branch, compare and export. Python workers persist independently of the DSH turn, expose durable state and confirm cancellation. Exports contain a model, recipe, report, pinned dependencies and standalone reproduction; data is excluded by default.

MLflow 3.18.0 and scikit-learn supply tracking, pipelines, training algorithms, metrics and model serialization. Optuna, DVC and RD-Agent were reviewed but are not runtime integrations in this release.

Validation: 13 Python cases and 3 Node cases; actual DSH tool executor and plugin-manager installation; seven real synthetic-data runs; two exported models independently retrained with matching metrics. See validation evidence for exact scope and clean-environment results.

An online DSH account-provider call returned ACCOUNT_SIGN_IN_REQUIRED before inference. Human intervention reduction and native-versus-plugin agent gains remain unmeasured. The release targets local single-user CSV classification/regression, not time-series or grouped evaluation, multi-tenant service, automatic tuning, or mid-training checkpoint resume.

Install the Python wheel with the supplied requirements.lock, then the DSH .tgz, or add this third-party index to DSH Discover:

https://raw.githubusercontent.com/shinerain1/dsh-ml-workbench/main/market/index.json

Configure absolute pythonExecutable and root paths as shown in the README. Only enable custom factories for reviewed local Python. This is an independent community plugin/index, not an official approval claim.