Skip to main content
Synthetic data is training data that a generator model writes from a few seed examples. A team uses it to train before it has real rows, or when it cannot train on the rows it has. monte dataset generate writes and filters the rows, and monte dataset add stores the rows that pass as a dataset. Synthetic data: seed examples, generate, filter by duplicates, judge model, and grader, review a sample, and add the rows Synthetic data: seed examples, generate, filter by duplicates, judge model, and grader, review a sample, and add the rows

Write the Seeds

Seeds are demonstration rows that show the task. The generator copies the faults of the seeds as well as their strengths, and it writes rows of the same shape as the seeds. To filter with a grader, each seed also holds the grader fields of that training environment, and the generator writes new grader fields for each row. When the environment’s tasks declare tools, as support-escalation-train does, each seed is a chat messages row with the assistant’s tool_calls, so the grader can score each tool call. On one line each:
A task without tools takes input and output seeds. For --kind preference, each seed is a preference row with prompt, chosen, and rejected. The judge and the grader score chosen. The default kind is sft. Seeds can come from the team or from public documents. Real requests still make the best eval tasks. Start with 20 to 50 seeds that cover the task’s categories.

Generate and Filter the Rows

The generator writes rows that vary the seeds in topic and wording. Three filters then drop rows:
  • Deduplication drops rows that repeat a seed or another row.
  • A judge model scores each row from 0 to 1 for correctness and agreement with the seeds.
  • A grader from a training environment scores each reply against the row’s grader fields. It checks that the reply agrees with its row’s own fields, not that the fields are correct. The sample review checks the fields.
A row passes when each score is at or above --min-score. The command writes the rows that pass to support-synth.jsonl, and the dropped rows, each with its reason, to support-synth.dropped.jsonl. It does not add a dataset. Dataset Commands lists every flag, including multi-turn rows and topic lists.

Generator and Judge Models

The generator and the judge are open-weight models. Each is a Hugging Face ID, such as Qwen/Qwen3-235B-A22B-Instruct-2507, or the name of a deployment, such as qwen3-235b. The judge comes from another model family, such as zai-org/GLM-4.7, because a model rates text in its own style too high. Both models run in Monte Cloud, or in the team’s AWS, GCP, or Azure account with BYOC. No seed row or generated row goes to a model provider. Security and Privacy lists the model calls that leave the cloud.

Review and Add

Read a sample of the kept rows and the dropped rows before the add. The dropped rows show whether --min-score is too strict or too loose. Then add the rows:
A generator that saw eval tasks can copy them. The add records each eval whose task text appears in the rows: see The Overlap Record.