Intent-scoped fabio skill for Fabric machine learning and data science: ML experiments (runs, metrics), ML models (registry, versions, endpoints, batch scoring), and anomaly detectors. Use to track experiments, register/version/serve models, score data, and configure anomaly detection. Triggers: "ml model", "ml experiment", "mlflow", "register model", "model version", "score model", "batch scoring", "model endpoint", "anomaly detection", "train model", "data science".
license
MIT
fabio-data-science — Data Science — ML experiments, models, versions, scoring, anomaly detection
Generated file — do not edit by hand. This intent-scoped sub-skill of the fabio skill is generated from fabio's command schema plus authored judgment. Regenerate with cargo test generate_subskills -- --ignored. For install, auth, output envelope, global flags, and agent-safety rules, see the root fabio skill.
Prefer runtime introspection. This index is a snapshot; the installed binary is always authoritative. Use fabio context agent --group <group> and fabio context describe <group> <command> for exact flags and output shapes.
When to use
Creating/listing ML experiments to organize training runs and metrics.
Reading MLflow run data of an experiment (list-runs with --filter/--order-by, get-run, get-metric-history) to compare runs and inspect logged metrics.
Registering ML models and managing versions (list-versions, get-version, activate-version, deactivate-version).
Serving/scoring: get-endpoint/update-endpoint, score and score-version for batch inference.
Creating and configuring anomaly detectors (get-definition/update-definition).
When NOT to use (route elsewhere)
Writing the training code itself (notebooks/Spark) -> use fabio-data-engineering.
The lakehouse feature tables the model reads/writes -> use fabio-lakehouse.
BI semantic models ('model' in a reporting sense) -> use fabio-bi (see disambiguate model).
Command index
Generated from fabio's command schema. For full flag details use fabio context agent --group <group> or fabio context describe <group> <command>.
fabio ml-experiment
Manage ML experiments (data science)
Command
Mutates
Description
fabio ml-experiment create
yes
Create a new ML experiment
fabio ml-experiment delete
yes
Delete an ML experiment
fabio ml-experiment get-metric-history
no
Get the logged history of a single metric across a run's steps
fabio ml-experiment get-run
no
Show details of a single run (info, parameters, metrics, tags)
fabio ml-experiment list
no
List ML experiments in a workspace
fabio ml-experiment list-runs
no
List runs in an experiment (MLflow tracking: parameters, metrics, status)
fabio ml-experiment show
no
Show details of an ML experiment
fabio ml-experiment update
yes
Update ML experiment properties (name and/or description)
fabio ml-model
Manage ML models (data science)
Command
Mutates
Description
fabio ml-model activate-version
yes
Activate a specific endpoint version
fabio ml-model create
yes
Create a new ML model
fabio ml-model deactivate-all-versions
yes
Deactivate all endpoint versions
fabio ml-model deactivate-version
yes
Deactivate a specific endpoint version
fabio ml-model delete
yes
Delete an ML model
fabio ml-model get-endpoint
no
Get the ML model serving endpoint configuration
fabio ml-model get-version
no
Get a specific endpoint version
fabio ml-model list
no
List ML models in a workspace
fabio ml-model list-versions
no
List endpoint versions
fabio ml-model score
no
Score against the ML model endpoint
fabio ml-model score-version
no
Score against a specific endpoint version
fabio ml-model show
no
Show details of an ML model
fabio ml-model update
yes
Update ML model properties (name and/or description)
fabio ml-model update-endpoint
yes
Update the ML model serving endpoint configuration
Create/select an ML experiment to track runs before registering a model from the best run.
Reference a specific model version for reproducible scoring (score-version), not just the active alias.
Author anomaly-detector logic via its definition (get-definition/update-definition).
PREFER
Experiment tracking (ml-experiment) over ad-hoc metric logging so runs are comparable.
Batch scoring via score/score-version over hand-rolled inference loops.
Runtime introspection (context agent --group ml-model|ml-experiment) for exact flags.
AVOID
Confusing an ML model with a BI semantic model — different groups (see disambiguate model).
Deactivating all model versions (deactivate-all-versions) without confirming nothing serves from them.
Scoring against an unpinned version when reproducibility matters.
Key gotchas
ML models are versioned: a model item holds multiple versions; activate-version sets the serving alias while list-versions/get-version address specific ones.
Scoring has two forms: score (active version) and score-version (a pinned version); endpoints are managed separately (get-endpoint/update-endpoint).
Anomaly detectors carry their logic in a definition part (base64) — edit via update-definition, not plain flags.
ml-experiment {list-runs,get-run,get-metric-history} read the per-workspace Fabric-hosted MLflow tracking server (the experiment item GUID is the MLflow experiment_id) — NOT the Fabric item API; --filter/--order-by take MLflow expressions.
Troubleshooting
Symptom
Fix
Scoring returns unexpected results after a retrain
The active version changed; pin with score-version , or confirm which version is active via list-versions.
Model endpoint not responding
Check get-endpoint; the serving endpoint is managed separately from version activation.
Runs are hard to compare
Group them under an ml-experiment and log metrics per run rather than ad-hoc.
Anomaly detector changes ignored
Edit its definition via update-definition (base64 part); metadata-only update does not change logic.
Safety
deactivate-all-versions / deactivate-version can take a served model offline — confirm nothing depends on it.
Deleting an ml-model removes all its versions — irreversible; confirm with the user.
Shared references
Cross-cutting operational guidance (the "common" layer) — consult the relevant topic before non-trivial work:
Reference
Covers
fabio context best-practices throttling
fabio transparently handles 429 (Too Many Requests) and gateway errors. Agents do NOT need to implement retry logic.
fabio context best-practices pagination
fabio handles pagination via --all (auto-fetch all pages), --continuation-token (resume), and --limit (truncate). Agents rarely need to paginate manually.
fabio context best-practices lro
Many Fabric operations are async (return 202). fabio polls them automatically. Use --wait for job operations.
