Build a leakage-aware training workflow with explicit data splits, measured resource budgets, held-out evaluation and traceable artifacts.
Write the target, prediction time, eligible population, decision use and a simple baseline. Identify repeated people, groups or time dependence before splitting rows. Record data version, label availability and the metric that reflects the intended task, including important error slices and excluded cases.
Choose random, stratified, grouped or chronological splitting according to the data-generating process. Use the allocation tool for whole-count planning only, then save actual row/group identities and inspect coverage. Fit preprocessing and feature selection within training boundaries; keep final evaluation data out of model and threshold choices.
Implement the complete preprocessing and estimator pipeline, then run a small end-to-end training and prediction pass. Select a compatible validation method; the K-fold tool counts ordinary folds and their remainders but cannot choose the best k or enforce group isolation. Save the exact validation partitions used for comparisons.
Inventory the actual model’s unique parameters, resident weight/state components and representative batch shapes. Use parameter formulas only for matching architectures and record separate GPU memory components and reserve. Measure complete representative training throughput and overhead before extrapolating runtime; parameter count alone does not determine memory or speed.
Record configuration, data and code versions, software/hardware, seeds and checkpoints. Compare candidates within the planned validation procedure, then evaluate the chosen pipeline on the reserved final data. For a numeric regression target, inspect actual/predicted pairs with R² and RMSE; use suitable different metrics for classification or other tasks.
Save preprocessing, model, target definition, feature schema and reference predictions together. Re-run inference from the saved artifact in a clean compatible environment and compare expected outputs. Report final metrics, slices, uncertainty and known limits; seeds alone do not guarantee identical results across platforms or library releases.
Write and download the task brief, measured results and handoff decisions as Markdown.
Allocate whole sample counts without losing rows; implement the appropriate split on the actual data separately.
Account for larger remainder folds and their corresponding training counts in an ordinary K-fold plan.
Check a matching dense, grouped Conv2D or explicitly specified decoder parameter count.
Build a per-device resident-memory ledger with explicit state components, budget and reserve.
Inspect a fixed-shape linear tensor-storage budget before testing real candidate batch sizes.
Estimate the entered training workload from measured sample throughput and separately entered overhead.
Evaluate actual/predicted numeric pairs with unweighted R² and RMSE, retaining undefined constant-target cases.
Loading your checklist…
Scikit-learn explains fitting preprocessing only on training data and keeping transformations consistent at evaluation and serving.
PyTorch notes that reproducibility is not guaranteed across releases or platforms; retain the environment and actual validation evidence alongside seeds.
Keep the following evidence with the actual version used for this task.
| Record | Include | Check |
|---|---|---|
| Data | Version, labels, split identities and features | No future or held-out information used in fitting |
| Run | Code, environment, shapes, seed and checkpoint | Repeat reference predictions |
| Evaluation | Metric, baseline, slices and final boundary | Separate tuning from final assessment |
Shared weights can appear in multiple modules without representing separate storage.
A complex model needs evidence of value relative to a useful reference, not just a successful training run.