Datasets

These commands add, generate, and remove datasets. Datasets explains the row kinds and the overlap record.

Command	Description
`monte dataset add`	Add a dataset as a new version
`monte dataset generate`	Generate rows from seed examples
`monte dataset list`	List datasets
`monte dataset show`	Show one dataset and check its rows
`monte dataset delete`	Remove a dataset from artifact storage

monte dataset add

monte dataset add <name> --from-file <path> --kind sft|preference [key flags] [anonymize flags] [--json]
monte dataset add <name> --from-hf <repo>[@<rev>] --file <path> --kind sft|preference [key flags] [anonymize flags] [--json]

Add one JSONL file as a new dataset version. The command checks every row, records the overlap with every eval environment, and pushes the rows to artifact storage.

Flag	Type	Default	Description
`<name>`	argument	required	Lowercase letters, digits, and `-`
`--from-file`	path		A local JSONL file
`--from-hf`	string		A Hugging Face dataset repo. `@<rev>` pins a branch, tag, or commit.
`--file`	path	required with `--from-hf`	The JSONL file inside the repo
`--kind`	string	required	`sft` for demonstrations, `preference` for preference pairs
`--input-key`	string	`input`	The prompt column of an `sft` row
`--output-key`	string	`output`	The answer column of an `sft` row
`--prompt-key`	string	`prompt`	The prompt column of a `preference` row
`--chosen-key`	string	`chosen`	The preferred answer of a `preference` row
`--rejected-key`	string	`rejected`	The rejected answer of a `preference` row
`--anonymize`	flag	off	Replace personal details before the rows are stored. Anonymization explains it.
`--entities`	list	every built-in type	With `--anonymize`: the types to find. NER types: `person`, `org`, `location`, `address`. Pattern types: `email`, `phone`, `ip`, `account`, `card`, `id`
`--replace`	`placeholder`, `pseudonym`, or `redact`	`placeholder`	With `--anonymize`: what a found value becomes
`--threshold`	number from 0 to 1	`0.5`	With `--anonymize`: the lowest NER confidence that counts as a find
`--keep`	path	none	With `--anonymize`: a text file of terms that are never replaced, one for each line
`--patterns`	path	none	With `--anonymize`: a YAML file that maps a custom type to a regular expression
`--preview`	integer	off	With `--anonymize`: print this many rows before and after, and store nothing
`--json`	flag	off	Print JSON on stdout

monte dataset add support-replies --from-file demos.jsonl --kind sft --anonymize

checked     1,240 rows (sft: messages)
anonymized  PERSON 2,118  EMAIL 342  PHONE 97  ORG 1,406  ACCOUNT 21
overlap     none against support-escalation@1, support-escalation-dev@2
added       support-replies@1  sha256 0be4a6f9...

Error	Cause
`E_REFUSAL`	A row does not match the kind. The message names the line.
`E_NOT_FOUND`	No file at the path, or no such repo, revision, or file on Hugging Face

monte dataset generate

monte dataset generate --seeds <path> --name <dataset> --model <model> [--kind sft|preference] [--rows <n>]
  [--turns <n>] [--topics <path>] [--dedup near|exact|off] [--judge <model>]
  [--grader <environment>] [--min-score <x>] [--out <path>] [--remote] [--json]

Generate rows from seed examples, and drop the rows that fail the filters. The rows that pass go to a JSONL file for review, and the dropped rows go to a second file with a reason for each. The command does not add a dataset. Synthetic Data explains each step. The generator and the judge are open-weight models. Each is a Hugging Face ID, such as Qwen/Qwen3-235B-A22B, or a deployment name, such as qwen3-235b.

Flag	Type	Default	Description
`--seeds`	path	required	A JSONL file of seed rows in the format of `--kind`. With `--grader`, each seed also holds that environment’s grader fields.
`--kind`	string	`sft`	`sft` writes `input` and `output` rows. `preference` writes `prompt`, `chosen`, and `rejected` rows from preference seeds.
`--name`	string	required	The name of the dataset to come. It also names the output files.
`--rows`	integer	`500`	The number of rows to generate, before the filters
`--model`	string	required	The generator: an open-weight model ID or a deployment name
`--turns`	integer	`1`	The exchanges in each row. Above 1, each row is a conversation in the `messages` format.
`--topics`	path		A text file of topics, one on each line. Without it, the generator takes topics from the seeds.
`--dedup`	string	`near`	`near` drops rows close to a seed or another row, `exact` drops only exact copies, `off` keeps every row
`--judge`	string		A judge that scores each row from 0 to 1: an open-weight model ID or a deployment name
`--grader`	string		A training environment whose grader scores each reply. An eval environment is refused.
`--min-score`	number	`0.5`	The lowest judge or grader score that passes
`--out`	path	`./<name>.jsonl`	The output file. The dropped rows go to the same path with `.dropped.jsonl`.
`--remote`	flag	off	Run the job on a GPU instance
`--json`	flag	off	Print JSON on stdout

monte dataset generate --seeds seeds.jsonl --name support-synth --rows 800 \
  --model Qwen/Qwen3-235B-A22B --judge Qwen/Qwen3-235B-A22B --min-score 0.7 \
  --grader support-escalation-train --remote

generated   800 rows
dropped     31 near copies, 42 below 0.7 on the judge, 18 below 0.7 on support-escalation-train@1
wrote       ./support-synth.jsonl (709 rows), ./support-synth.dropped.jsonl (91 rows)

Error	Cause
`E_REFUSAL`	A seed row does not match the format of `--kind`, or `--grader` names an eval environment

monte dataset list

monte dataset list [--stored] [--json]

List the dataset versions on this machine.

Flag	Type	Default	Description
`--stored`	flag	off	List the datasets in artifact storage
`--json`	flag	off	Print JSON on stdout

monte dataset list

NAME              KIND         VERSION   ROWS    OVERLAP
support-replies   sft          1         1,240   none
support-synth     sft          1         709     support-escalation-dev@2 (3 rows)

monte dataset show

monte dataset show <name>[@<version>] [--json]

Show one dataset: its kind, origin, row count, columns, anonymization counts, and overlap record. The command also checks the rows against their hash, and fetches the dataset from artifact storage if this machine does not hold it.

Flag	Type	Default	Description
`<name>[@<version>]`	argument	required	The dataset. A bare name means the newest version.
`--json`	flag	off	Print JSON on stdout

monte dataset show support-replies

dataset     support-replies@1 (sft)  rows 1,240  sha256 0be4a6f9... verified
anonymized  PERSON 2,118  EMAIL 342  PHONE 97  ORG 1,406  ACCOUNT 21
overlap     none
used by     support-triage/train-07, support-triage/train-09

monte dataset delete

monte dataset delete <name>[@<version>] [--yes] [--json]

Remove a dataset from artifact storage and this machine. A bare name removes every version. Run history keeps the record that a run trained on the dataset, but not the rows.

A delete does not remove what a model learned from the rows. Delete a Dataset explains what else to remove.

Flag	Type	Default	Description
`<name>[@<version>]`	argument	required	The dataset, or one version of it
`--yes`	flag	off	Skip the prompt
`--json`	flag	off	Print JSON on stdout

monte dataset delete support-replies

This removes 1 version of support-replies. 2 training runs used it: support-triage/train-07, support-triage/train-09.
delete? [y/N] y
Deleted support-replies@1 from artifact storage and this machine.

Commands

MCP

monte dataset add

monte dataset generate

monte dataset list

monte dataset show

monte dataset delete

​monte dataset add

​monte dataset generate

​monte dataset list

​monte dataset show

​monte dataset delete

monte dataset add

monte dataset generate

monte dataset list

monte dataset show

monte dataset delete