Skip to content

Data Preparation

Solver code reads a single workspace-relative root: workspace/dataset/. Raw task data is treated as read-only. Preparation may create metadata and derived splits inside the workspace, while large unchanged media directories are linked rather than copied.


Mode Behavior Output
deepbuilder prep Creates an empty workspace dataset directory, loads the repository skill registry, and runs the data-preparation ScienceAgent. Prepared dataset/, optional split manifest, and a human-readable report.
deepbuilder prep --link-only Skips the agent and symlinks every top-level input entry into the flat dataset view. A fast, non-semantic workspace view.
deepbuilder repl Uses the same flat-symlink fast path when the workspace has no dataset view yet. Dataset access for an interactive agent session.

Inspect Inputread-only task data
List Skillscategory=data_processing
Read Guidancedata_prep + matching skill
Build Datasetflat files & safe symlinks
Audit Splitmanifest & report

solver/data_prep.py requires the agent to list data-preparation skills, read data_prep, inspect the input root, and write only within the workspace. It uses the standard ScienceAgent with a stable system prompt, bash-write tools, repository skills, and a dedicated memory directory.


  • Group-aware splitting: Prefer group-aware splitting over row-level random splitting when stable patient, study, source, sequence, or paired-item identifiers exist.
  • Task-like validation: Prefer task-like validation when test examples are constructed differently from raw training rows.
  • Stratification: Use stratification or coverage balancing for imbalanced, multilabel, medical, detection, segmentation, and sparse-positive tasks when feasible.
  • Strict isolation: Keep held-out groups and rows completely outside the remaining training files.
  • No data leakage: Do not use private answers, test labels, or final-test keys for tuning.
  • Honest reporting: If a safe split cannot be inferred, preserve the usable data view and document the limitation instead of manufacturing a misleading holdout.

Artifact Contract
dataset/train* Training-side records and labels after any holdout has been removed.
dataset/validation* Optional local validation records in the schema expected by the task.
dataset/mask_validation_label.* Optional hidden labels when validation features must not expose the target.
dataset/split_manifest.json Machine-readable seed, strategy, group key, target columns, counts, distributions, and validity notes.
dataset/split_report.md Short explanation of the resulting layout, validation policy, and known limitations.

The repository currently provides data_prep, image-task guidance, BSON image loading, recommendation time splitting, tabular preprocessing speedups, and self-evaluation guidance under .deepbuilder/skills/data_processing/. New dataset knowledge should normally be added as a skill rather than embedded as task-specific branches in the runtime.