Runs and Results

Runs

Each pipeline execution is a run, also called a session, and has a session ID. The dashboard's Runs tab lists all of your own runs, newest first:

• training runs;
• inference and retraining runs;
• runs started by automations.

Runs that other people share with you appear on the Collaboration tab instead.

At the top of the tab are counts of total, completed, failed and cancelled runs. You can search by name, file or session ID, filter by status, and filter by tag. Each row shows:

• the run name and the first part of its session ID, with a button that copies the full ID;
• the run type: Pipeline, Inference or Retraining;
• the input file and its row count, column count and size;
• the start time;
• the status.

Row actions

Action	What it does
Rename	Hover over the name and click the pencil. Names can be up to 255 characters. API: PUT /api/v1/sessions/<id>/name with {"session_name": "..."}.
Tags and note	Up to 10 tags of up to 32 characters each, and a private note of up to 500 characters. API: PUT /api/v1/sessions/<id>/meta.
Zip	Downloads every output as autodata_<first 8 characters of the ID>.zip. Available on completed runs.
Clone config	Opens the Pipeline tab filled in with this run's configuration, so you can start a new run on a new file. The link form is /dashboard/pipeline?clone=<session_id>.
Retry	Shown on failed runs. It resubmits the run from its last finished stage, using the same session ID. The retry needs the original upload, and uploads are removed when a run ends. So in practice it works only after an interruption such as a server restart. Otherwise, clone the configuration and upload the file again. API: POST /api/v1/sessions/<id>/retry.
Show files	Lists the run's output files, each with a download button.
Compare	Tick two runs, then click Compare. You see rows and columns in and out, duration, output files and warnings side by side, with differences highlighted.

Pipeline templates

The Templates menu on the Pipeline tab can save the current configuration under a name of up to 80 characters. You can then apply, rename or delete saved templates. Each account can hold up to 50 templates, and each template's configuration can be up to 64 KB. Templates include media settings. They do not include a run mode.

API: GET and POST /api/pipeline-templates; PUT and DELETE /api/pipeline-templates/<id>.

Starting, following and cancelling a run

• Confirmation. Before a dashboard run starts, a dialog shows a recap and the exact price of the uploaded file with your settings.
  • Time is shown for each speed mode, relative to Rapid.
  • The price is the same in every speed mode. Confirm $X and start starts the run, and you are charged exactly that amount. If the file or a setting that affects the price changed since, the run is not started and the new price is shown for you to confirm. If the price is above your balance you are warned, but the run can still start.
  • For every setting you configured that the chosen speed mode would change, you pick Keep my settings or Use <mode>.
  • The API equivalent is POST /api/v1/estimate with the pricing_profile_id from validation. It never charges.
• Waiting for approval. Scheduled runs, triggers and folder listeners can have a Maximum price per run ($). A run whose price is above it is held instead of running and listed under Waiting for approval on the Runs tab, with its price and the limit. Approve starts it at that price, and it is charged exactly that; Discard drops it and nothing is charged. You are notified by a job.held webhook event and, if you receive failure emails, by email. API: GET /api/pending-approvals, POST /api/pending-approvals/{id}/approve, POST /api/pending-approvals/{id}/discard (409 if already decided, 410 if the held data is gone).
• Duplicate protection. Send an Idempotency-Key header, or an idempotencyKey form field, with /process-data. If a later request repeats that key, you get back the existing session ID instead of a second run.
• Progress. The Pipeline tab shows a stage timeline with a duration for each stage, a progress bar for long stages, and the warnings raised so far. A sidebar chip, Run in progress · Step N/M, follows the run from any tab. Keep the dashboard open until a run finishes the first time, because a dashboard run is added to your history when its result is first fetched.
• Cancel. Stops the run, which is then recorded as Cancelled. Files from stages that already finished may remain. API: POST /api/v1/cancel/<session_id>.

Output files

You choose which files are written under Output Preferences. These are saved per account, and each run can override them. A stage's file is written only when its preference is on and the stage actually ran.

File (step)	Default	Contents
anormaly_fixed_output	on	Data after anomaly repair
dtc_output	on	Data after type conversion and encoding
mdh_output	on	Data after missing values have been handled
cds_output	on	Scaled data
dsm_output	on	Data after feature selection. Contains all real rows, fully processed.
dsg_output	on	The processed data plus synthetic rows, up to your requested row count. Real rows come first.
multimodal_output	on	The final merged frame
model_ready	on	Flattened, all-numeric data with no empty cells, plus model_ready_manifest.json. See below.
compleated_dataset	off	The Data Completion and Verification result. Always CSV.
pipeline_report.pdf	on	The PDF pipeline report
quality_report.json	always	A baseline model score on the real, processed rows
result_evaluation_report.json and .csv	off	A benchmark of several standard models. Written only when the Result Evaluator is on. Included in ZIP downloads.
step_metrics.json and .csv	always	Timing and row counts for each stage

Which file to evaluate on

dsg_output is not an evaluation set. If the row count you requested is larger than your data, most of its rows can be synthetic, and a model scored on it will look better than it is.

Score on one of these instead:

• dsm_output, which contains the real rows, fully processed;
• the first real_rows rows of dsg_output;
• the model-ready test split, which is drawn from real rows only by default.

Every result reports dsg_provenance, which gives the real and synthetic row counts.

Model-ready export

Setting	Values	Default
model_ready_split	none, xy (X and y files), train_test, xy_train_test (X_train, X_test, y_train and y_test)	none
model_ready_test_size	0.05 to 0.5	0.2
model_ready_real_test_only	Draw the test split from real rows only	on

Columns that cannot be made numeric are written to model_ready_excluded_columns, in the same row order as the export. If a split cannot be made, the export falls back to a simpler layout and adds a warning.

File names

output_name_pattern controls how data files and model-ready files are named:

• prefixed (the default) gives [input]_[YYYYMMDD_HHMMSS]_[step].[ext], for example sales_20260922_013000_dsg_output.parquet. Automated runs leave out the input part.
• plain gives just dsg_output.parquet.

Reports, manifests and inference or retraining outputs always keep their fixed names.

Export format

output_format can be csv (the default), parquet, xlsx, json, jsonl, ndjson, feather or orc. The format never makes a run fail:

• A file that is too large for Excel, which allows 1,048,576 rows and 16,384 columns, is written as CSV instead.
• A file that cannot be written in the chosen format is also written as CSV instead.
• Columns of mixed type in Parquet, Feather or ORC are written as text.

Every such change is listed in output_format_warnings. Inference and retraining outputs are always CSV.

Downloading

Route	Returns
GET /download/<session_id>/<filename>	One file
GET /api/sessions/<session_id>/files	The list of files, with their sizes and download URLs
GET /api/v1/sessions/<session_id>/download-all	A ZIP of every output. Your input file and empty files are left out.
POST /api/v1/download-archive/<session_id>	A ZIP of selected files. Send {"files": ["dsg_output", "model_ready"]}; names can be full file names, stems or step names.

Outputs are always stored on disk. The older in-memory output routes under /api/v1/sessions/<id>/outputs are kept for compatibility, but current runs do not fill them.

How long files are kept

What	Kept for
Your uploaded input file	Removed when the run ends
Output data files, the PDF, model-ready files and most reports	7 days
The run's stored configuration and transforms, which inference and retraining need, plus uploaded images and audio	1 year
Run history, names, tags and notes	Not removed

Download what you need within the week, or set the run to write its results to a database automatically. Self-hosted installations can change these periods.

Statuses and errors

Status	Meaning
Processing	The run is in progress. A run stuck in this state for more than 2 hours is marked failed.
Completed	Every enabled stage finished.
Failed	A stage stopped the run. The result carries error_info with a code, stage, column, message and hint. The code is one of TOO_FEW_ROWS, TARGET_INVALID, COLUMN_NOT_FOUND, OUT_OF_MEMORY, NO_DATA, WORKER_DIED, TIMEOUT, DATA_TYPE_CONFLICT, FILE_UNREADABLE or PROCESSING_FAILED.
Cancelled	You stopped the run.

To list runs, call GET /api/v1/sessions. It accepts these filters: q, status, tag, date_from and date_to. Each run in the response includes is_retrainable, which tells you whether its transforms are still stored. To get one run's details, call GET /api/v1/sessions/<id>. For the configuration used to clone a run, call GET /api/v1/sessions/<id>/config.