Deepchecks Repo Skill

Use this repo skill when a task involves the Python package Deepchecks for validating ML data, train/test splits, model behavior, drift, weak segments, result reports, or CI gates across tabular, NLP, and computer-vision workflows.

Install and import route

Start with the smallest install that matches the workflow:

pip install deepchecks
pip install "deepchecks[nlp]"              # NLP TextData, text checks, transformers-backed helpers
pip install "deepchecks[nlp-properties]"   # optional heavier text-property calculators
pip install "deepchecks[vision]"           # vision helpers; also install compatible torch/torchvision when needed

Then run the safe bundled diagnostic when imports or optional extras are uncertain:

python scripts/check_deepchecks_install.py --include-nlp --include-vision

Read package overview for component boundaries, optional extras, Python-version notes, no-CLI expectations, and latest-version-check behavior. Read root troubleshooting for install/import/display/optional dependency failures. Read repository provenance before deciding whether this skill is stale for a checkout.

Route by user task

User asks for...	Load this sub-skill
pandas/numpy data validation, Dataset, tabular data integrity, train-test split validation, model evaluation, tabular checks/suites, custom scorers, supplied predictions or probabilities	tabular-validation
raw or tokenized text, TextData, text classification/token classification labels, text metadata/properties/embeddings, NLP suites/checks, tokenizer/model-download avoidance	nlp-validation
images, object detection, semantic segmentation, VisionData, BatchOutputFormat, PyTorch/TensorFlow/custom loaders, vision checks/suites, torch/torchvision or CUDA questions	vision-validation
CheckResult/SuiteResult display, HTML reports, JSON serialization/recovery, pytest assertions, CI gates, GitHub Actions/Airflow/H2O/Hugging Face adapter patterns	results-and-integrations

Operating rules for future agents

1. Choose the data modality first; each modality uses a different Deepchecks data container and validation context.
2. Prefer built-in suites for first-pass coverage, then switch to individual checks when the user names a specific issue or threshold.
3. Use explicit metadata (label, categorical columns, task type, label map, properties, embeddings, prediction/probability arrays) instead of relying on inference when the user has schema knowledge.
4. Avoid downloads in diagnostic or CI paths. Prefer tiny local fixtures, precomputed predictions, precomputed embeddings/properties, and bundled smoke scripts.
5. Save reports before failing CI. Route result persistence and gates to results-and-integrations instead of duplicating export logic in modality sub-skills.
6. Treat visible GPUs as optional unless the user specifically needs GPU-backed model execution. Deepchecks validation workflows can often be prepared and diagnosed with CPU-capable dependencies.
7. Do not tell users to run original repository examples or tests as part of package use. Use the bundled scripts and distilled references in this skill tree.

Bundled diagnostics

• scripts/check_deepchecks_install.py: check base/NLP/vision imports and optional torch CUDA status without running suites or downloading data.
• Each modality sub-skill has its own smoke helper for tiny local Dataset, TextData, or VisionData construction.

When not to use this skill

• The task is about Deepchecks Monitoring self-hosted deployment or managed SaaS operations rather than the open-source testing package APIs.
• The task is generic ML training, model serving, dataset labeling, or experiment tracking without Deepchecks checks, suites, or result artifacts.
• The task is maintainer-only release, benchmark, docs-build, or repository CI infrastructure work unless the user explicitly asks to modify this Deepchecks checkout.