Skip to main content

AutoML

Create, manage, and monitor automated machine learning jobs. AutoML automatically selects the best model architecture, hyperparameters, and feature engineering for your dataset and prediction task.

All endpoints require authentication via X-API-Key header and the appropriate scope.

AutoMLJob Object​

{
  "_id": "Xk3pQ9rT2mB7nW4sA",
  "jobId": "f22a9764e2ed836b9799a50b",
  "name": "Churn Prediction Model",
  "problemType": "tabular",
  "predictorType": "auto",
  "status": "training",
  "progress": 45,
  "owner": "user_456",
  "organizationId": "org_xyz",
  "trainingConfig": { "preset": "medium_quality", "timeLimit": 600, "metric": "accuracy" },
  "hardware": { "cpuCount": 4, "memoryGb": 16, "gpuCount": 0, "diskGb": 20 },
  "environmentId": "custom",
  "datasetInfo": {
    "name": "Customer Data Q4 2024",
    "size": 45000,
    "rows": 0,
    "columns": 0,
    "targetColumn": "churned",
    "featureColumns": ["usage_days", "monthly_spend", "support_tickets", "plan_type"],
    "volumeId": "vol_customers_001",
    "path": "data/customers.csv"
  },
  "outputVolumeId": "vol_models_001",
  "cost": { "estimatedTotal": 10.5, "currentCost": 0 },
  "metrics": { "trainingTimeHours": 0 },
  "createdAt": "2025-02-01T09:55:00Z",
  "updatedAt": "2025-02-01T12:30:00Z"
}

jobId is the job's ID everywhere in the API; :id in a path also accepts _id. status is one of pending, preparing, running, training, evaluating, stopping, completed, failed, cancelled, with status_message describing progress. A failed job carries error (and failure_reason). A completed job records its scores in metrics and its trained models in models_info (best_model, leaderboard and, for tabular jobs, the trained predictor's details). Its scores (metrics.val_score, metrics.test_score, the leaderboard's) are AutoGluon's in models_info.eval_metric, where higher is always better: an error metric (RMSE, MAE, MASE, log loss and the like) is reported negative, and its value is the score with the sign flipped (a val_score of -19990.88 in root_mean_squared_error is an RMSE of 19,990.88). Each model's validation score is as it serves: a binary model calibrated to a decision threshold is scored with it, as its test score and results are. metrics.feature_importance holds each feature's permutation importance as a share of the total, computed on the rows metrics.feature_importance_rows names ({"data": "validation" | "test", "rows": n}), which the model was not trained on. datasetInfo holds volumeId and path when the job trains on a volume, or s3_path when it trains on an uploaded file. organizationId is set when the owner belongs to an organization.

GET /api/v1/mlops/automl-jobs​

List all AutoML jobs accessible to the authenticated user.

Scope: ml-workbench:read

Query Parameters

ParameterTypeRequiredDescription
statusstringNoFilter by status: pending, preparing, running, training, evaluating, stopping, completed, failed, cancelled
qstringNoSearch by job name or description
limitintegerNoNumber of results to return (default: 50, max: 200)
cursorstringNometa.nextCursor of the previous page; omit for the first page
sortstringNoSort field; prefix with - for descending (default: -createdAt)

Response 200 OK

{
  "data": [
    {
      "_id": "Xk3pQ9rT2mB7nW4sA",
      "jobId": "f22a9764e2ed836b9799a50b",
      "name": "Churn Prediction Model",
      "problemType": "tabular",
      "predictorType": "auto",
      "status": "training",
      "progress": 45,
      "owner": "user_456",
      "organizationId": "org_xyz",
      "trainingConfig": { "preset": "medium_quality", "timeLimit": 600, "metric": "accuracy" },
      "datasetInfo": { /* see the AutoMLJob object */ },
      "createdAt": "2025-02-01T09:55:00Z",
      "updatedAt": "2025-02-01T12:30:00Z"
    }
  ],
  "meta": {
    "total": 10,
    "limit": 50,
    "nextCursor": null,
    "requestId": "req_abc123"
  }
}

POST /api/v1/mlops/automl-jobs​

Create a new AutoML job.

Scope: ml-workbench:write

Datasets are project volumes (uploaded through the UI or the volume data endpoints). Reference one by its volume id (from GET /mlops/automl-datasets); a volume is a directory, so when it holds more than one data file, pass datasetFile with the file key (discover files with GET /mlops/automl-datasets/:id/files). You may also pass an s3:// path or a URL directly as dataset.

A job can also train straight from one of your data sources (dataSourceId) or add-ons (addonId), with read naming what to read. Nothing is copied first: the job receives the source's connection through STRONGLY_SERVICES, as an app does, and reads every row read-only. The source is refused with 400 when it does not exist, cannot be read as read says, or is of a type a job cannot read (API keys, Memcached, RabbitMQ, SQS, MQTT, Marqo), and with 403 when you may not use it. A target column the platform cannot list (a query, a document store) is checked when the job reads: the job fails naming the columns it found.

{
  "name": "Churn from Postgres",
  "dataSourceId": "ds_sales_pg",
  "read": { "table": "churn", "schema": "public" },
  "targetColumn": "churned",
  "hardware": { "cpuCount": 2, "memoryGb": 8, "diskGb": 20 },
  "outputVolumeId": "vol_models_001"
}
Source typeread
SQL (Postgres and compatible, MySQL and compatible, SQL Server, Oracle, Snowflake, BigQuery, Redshift, ClickHouse, ...){ "table", "schema"? } or { "query" } (read-only)
MongoDB{ "collection", "filter"? } or { "collection", "pipeline" } (no $out/$merge)
Elasticsearch, CouchDB{ "collection" } (an index or database), with an optional filter
DynamoDB, Firestore{ "collection" } (a table or collection)
Couchbase, ArangoDB, SurrealDB{ "collection" } or { "query" } (a read-only SQL++, AQL or SurrealQL SELECT)
Neo4j, Neptune{ "query" } (Cypher or openCypher, read-only)
TigerGraph{ "collection" } (a vertex type)
Milvus, Pinecone, Weaviate, Qdrant, Chroma, Vespa{ "collection" } (Pinecone also namespace; Vespa namespace/documenttype). A vector becomes one column per dimension
Google Sheets, Airtable, Baserow, NocoDB, SeaTable, Grist{ "sheet" } (Google Sheets: optional, the first worksheet)
S3, MinIO, GCS, Azure Blob{ "object_key" }: a CSV, TSV, Parquet or JSON file, or a prefix ending in /
Redis{ "key_pattern" }: one row per key (a hash or a JSON value)
Kafka, Pulsar{ "topic" }: every message from the first to the last when the job starts; nothing is consumed

The trained model is written to a shared volume: pass exactly one of outputVolumeId (an existing shared volume) and outputVolumeName (a new one the platform creates). Otherwise the request is refused with 400 validation-error.

Request Body

{
  "name": "Churn Prediction Model",
  "dataset": "vol_customers_001",
  "datasetFile": "data/customers.csv",
  "targetColumn": "churned",
  "featureColumns": ["usage_days", "monthly_spend", "support_tickets", "plan_type"],
  "problemType": "tabular",
  "predictorType": "auto",
  "preset": "medium_quality",
  "timeLimit": 600,
  "metric": "accuracy",
  "hardware": { "cpuCount": 4, "memoryGb": 16, "diskGb": 20, "gpuCount": 0 },
  "outputVolumeId": "vol_models_001",
  "environmentId": "custom",
  "advanced": { "excludeModels": ["KNN"] }
}
FieldTypeRequiredDescription
namestringYesHuman-readable job name
datasetstringOne sourceProject dataset volume id (from GET /mlops/automl-datasets), an s3:// path, or a URL
dataSourceIdstringOne sourceA data source to train from, with read
addonIdstringOne sourceAn add-on to train from, with read
readobjectWith a data source or add-onWhat to read (see the table above)
targetColumnstringYesName of the column to predict
datasetFilestringNoFile key within the volume, when it holds more than one data file
featureColumnsstring[]NoSpecific columns to use as features. All non-target columns used if omitted
problemTypestringNoData modality: tabular (default), multimodal, timeseries
predictorTypestringNoauto (default, AutoGluon infers), BinaryClassifier, MultiClassifier, Regressor
presetstringNoAutoGluon quality preset: best_quality, high_quality, good_quality (tabular), medium_quality (default), optimize_for_deployment (tabular)
timeLimitintegerNoMaximum training time in seconds (default: 600)
metricstringNoOptimization metric (e.g. accuracy, f1, rmse). Default: accuracy
hardwareobjectYesPod sizing: { cpu_count, memory_gb, disk_gb }, with optional gpuCount, gpu_type (e.g. nvidia-t4, nvidia-a10g) useSpot (true trains on cheaper spot capacity, which can be reclaimed) and spotFallback (with useSpot, true, the default, falls back to on-demand when no spot is available; false waits for spot)
outputVolumeIdstringOne of the twoThe shared volume the trained model is written to
outputVolumeNamestringOne of the twoThe name of a new shared volume to create for the trained model
environmentIdstringNoTraining environment: custom (default) or an existing environment id
advancedobjectNoAdvanced settings, by section (see Advanced settings); an omitted section keeps its defaults

Advanced settings​

The sections of advanced, as the create page's Advanced step sets them. A job records what it ran with as advanced_config, and its page shows it on the Configuration tab.

SectionForKeys
model_selectionTabularenable_gbm, enable_cat, enable_xgb, enable_rf, enable_xt, enable_knn, enable_lr, enable_nn_torch, enable_fastai, enable_tabpfn, enable_ft_transformer, enable_automm (booleans: the model families to train); included_models, excluded_models (AutoGluon model names)
ensembleTabularnum_bag_folds, num_bag_sets, save_bag_folds, num_stack_levels, dynamic_stacking, fit_weighted_ensemble, fit_full_last_level_weighted_ensemble, use_orig_features, refit_full, set_best_to_refit_full, keep_only_best, save_space
data_handlingTabularfeature_generator (auto or none), enable_text_ngram_features, enable_text_special_features, enable_raw_text_features, enable_vision_features, feature_prune, feature_prune_threshold, sample_weight_column
trainingTabularinfer_limit (seconds per row), infer_limit_batch_size, calibrate, calibrate_decision_threshold, fit_strategy (sequential or parallel), verbosity, learning_curves, raise_on_model_failure
hpoTabularenabled, strategy (auto, random or bayesian), num_trials, time_limit_per_trial (seconds), scheduler (local or fifo)
model_hyperparametersTabularPer family: gbm, cat, xgb, rf, nn_torch, fastai (each family's AutoGluon hyperparameters). Values given are fixed for every model of the family, and tuning searches only the others; omit the section to train with AutoGluon's defaults for the preset, which hpo searches
customTabularhyperparameters_json, ag_args_json, ag_args_fit_json, ag_args_ensemble_json (JSON strings)
timeseriesTime seriesprediction_length, freq (e.g. D, H, W, M), id_column, timestamp_column, known_covariates, static_features, eval_metric_seasonal_period, quantile_levels
multimodalMultimodaltext_columns, image_columns, backbone, max_text_length, image_size, augmentation, use_ensemble, ensemble_size, ensemble_mode (one_shot or sequential), clean_ckpts

Response 201 Created

{
  "data": {
    "_id": "Xk3pQ9rT2mB7nW4sA",
    "jobId": "f22a9764e2ed836b9799a50b",
    "name": "Churn Prediction Model",
    "problemType": "tabular",
    "predictorType": "auto",
    "status": "pending",
    "progress": 0,
    "owner": "user_456",
    "organizationId": "org_xyz",
    "trainingConfig": { "preset": "medium_quality", "timeLimit": 600, "metric": "accuracy" },
    "hardware": { "cpuCount": 4, "memoryGb": 16, "gpuCount": 0, "diskGb": 20 },
    "environmentId": "custom",
    "createdAt": "2025-02-01T09:55:00Z",
    "updatedAt": "2025-02-01T09:55:00Z"
  },
  "meta": {
    "requestId": "req_abc123"
  }
}

GET /api/v1/mlops/automl-jobs/:id​

Get a single AutoML job by ID.

Scope: ml-workbench:read

Path Parameters

ParameterTypeRequiredDescription
idstringYesAutoML job ID

Response 200 OK

Returns the full AutoMLJob object.

DELETE /api/v1/mlops/automl-jobs/:id​

Delete an AutoML job. Running jobs must be stopped before deletion.

Scope: ml-workbench:write

Path Parameters

ParameterTypeRequiredDescription
idstringYesAutoML job ID

Response 204 No Content

POST /api/v1/mlops/automl-jobs/:id/stop​

Stop a running AutoML job. The job will be marked as stopped and training will be terminated.

Scope: ml-workbench:write

Path Parameters

ParameterTypeRequiredDescription
idstringYesAutoML job ID

Response 200 OK​

The job, as GET /mlops/automl-jobs/:id shows it, status cancelled.

POST /api/v1/mlops/automl-jobs/:id/deploy​

Deploy the best model from a completed AutoML job. Makes the model available for inference via the model registry.

Scope: ml-workbench:write

Path Parameters

ParameterTypeRequiredDescription
idstringYesAutoML job ID

Response 200 OK

{
  "data": {
    "success": true,
    "modelId": "model_xyz789",
    "modelName": "Churn Prediction Model",
    "version": 1,
    "deploymentStatus": "building",
    "message": "Best model registered and deployment started",
    "baselineJobId": "bjob_123"
  },
  "meta": {
    "requestId": "req_abc123"
  }
}

When the job wrote no reference rows for the model's drift baseline, baseline_note (why no baseline was built) is returned instead of baselineJobId.

GET /api/v1/mlops/automl-jobs/:id/logs​

Retrieve an AutoML job's training log: the trainer pod's while it exists, then the tail the job's record keeps. With no query parameters, the whole log.

Scope: ml-workbench:read

Path Parameters

ParameterTypeRequiredDescription
idstringYesAutoML job ID

Query Parameters

ParameterTypeRequiredDescription
linesintegerNoOnly the last N lines
sincestringNoOnly the lines written at or after this ISO 8601 time

Response 200 OK

{
  "data": {
    "logs": "Beginning AutoGluon training ...\nFitting model: LightGBM ...\n"
  },
  "meta": {
    "requestId": "req_abc123"
  }
}

logs is the training log as one text block.

GET /api/v1/mlops/automl-jobs/stats​

Get aggregate statistics for the AutoML jobs you can see.

averageAccuracy averages the validation accuracy of the completed jobs optimized for accuracy, and is null when there are none (another metric's score is not an accuracy). What AutoML jobs cost is shown in FinOps only.

Scope: ml-workbench:read

Response 200 OK

{
  "data": {
    "totalJobs": 10,
    "activeJobs": 1,
    "completedJobs": 6,
    "failedJobs": 2,
    "averageAccuracy": 0.87,
    "totalModels": 42
  },
  "meta": {
    "requestId": "req_abc123"
  }
}

GET /api/v1/mlops/automl-datasets​

List the shared dataset volumes the authenticated user can train on, most recently updated first. A volume is a directory; use It pages with limit and cursor (Pagination). GET /mlops/automl-datasets/:id/files to list the files inside it.

Scope: ml-workbench:read

Response 200 OK

{
  "data": {
    "datasets": [
      {
        "_id": "vol_customers_001",
        "name": "Customer Data Q4 2024",
        "description": "",
        "scope": "shared",
        "dataFileCount": 1,
        "hasData": true,
        "isShared": true,
        "owner": "user_456",
        "updatedAt": "2025-02-01T09:00:00Z",
        "accessLevel": "owner"
      }
    ]
  },
  "meta": { "total": 1, "limit": 50, "nextCursor": null, "requestId": "req_abc123" }
}

dataFileCount counts the training files (.csv, .tsv, .json, .parquet, .data, .txt, .xlsx, .xls) in the volume's data. accessLevel is owner or shared.

POST /api/v1/mlops/automl-datasets/uploads​

Get a presigned URL to upload a local training dataset file. PUT the file bytes to the returned uploadUrl, then reference the returned s3Path as the dataset in POST /mlops/automl-jobs. This is the programmatic equivalent of the UI's dataset uploader.

Scope: ml-workbench:write

Request Body

FieldTypeRequiredDescription
filenamestringYesDataset file name (e.g. titanic.csv)
contentTypestringNoMIME type (default text/csv)

Response 201 Created

{
  "data": {
    "datasetId": "f22a9764e2ed836b9799a50b",
    "uploadUrl": "https://s3.amazonaws.com/strongly-dev/mlops/automl/jobs/.../input/titanic.csv?X-Amz-...",
    "s3Path": "mlops/automl/jobs/f22a9764e2ed836b9799a50b/input/titanic.csv",
    "bucket": "strongly-dev",
    "expiresInSeconds": 3600
  },
  "meta": {
    "requestId": "req_abc123"
  }
}

Then PUT the raw file bytes to uploadUrl and create the job:

curl -X PUT "$uploadUrl" -H "Content-Type: text/csv" --data-binary @titanic.csv
# then POST /mlops/automl-jobs with { "dataset": "<s3Path>", "targetColumn": "Survived", ... }

GET /api/v1/mlops/automl-datasets/:id/files​

List the data files inside a project dataset volume, so you can pick which one to train on (pass its key as datasetFile to POST /mlops/automl-jobs).

Scope: ml-workbench:read

Path Parameters

ParameterTypeRequiredDescription
idstringYesDataset volume id (from GET /mlops/automl-datasets)

Response 200 OK

{
  "data": {
    "volumeId": "vol_customers_001",
    "dataVersion": null,
    "files": [
      { "key": "data/customers.csv", "name": "customers.csv", "size": 45000, "lastModified": "2025-02-01T09:00:00Z", "isDirectory": false }
    ]
  },
  "meta": {
    "requestId": "req_abc123"
  }
}

Only training files (.csv, .tsv, .json, .parquet, .data, .txt, .xlsx, .xls) are listed, sorted by name.

GET /api/v1/mlops/automl-datasets/:id/columns​

Get the column headers of a dataset file in a project volume (to choose the target and feature columns). Reads only the header row.

Scope: ml-workbench:read

Path Parameters

ParameterTypeRequiredDescription
idstringYesDataset volume id

Query Parameters

ParameterTypeRequiredDescription
filestringNoFile key within the volume. Optional when the volume has exactly one data file

Response 200 OK

{
  "data": {
    "volumeId": "vol_customers_001",
    "dataVersion": null,
    "file": "projects/proj_1/volumes/vol_customers_001/customers.csv",
    "columns": ["usage_days", "monthly_spend", "support_tickets", "plan_type", "churned"]
  },
  "meta": {
    "requestId": "req_abc123"
  }
}

GET /api/v1/mlops/automl-jobs/:id/status​

A job's status as the trainer reports it: status, progress, statusMessage, error, failureReason, createdAt, finishedAt, updatedAt, and while it runs k8s (the pod's live diagnostics: scheduling, image pulls, restarts). Lighter than the whole job.

Scope: ml-workbench:read

ParameterTypeRequiredDescription
idstringYesAutoML job id

Response 200 OK

GET /api/v1/mlops/automl-jobs/:id/metrics​

A job's training results: metrics, the leaderboard of the models it trained, the bestModel and trialCount.

Scope: ml-workbench:read

ParameterTypeRequiredDescription
idstringYesAutoML job id

Response 200 OK