JSONL Datasets

Each line of a JSONL dataset is one task: the prompt the agent sees, plus anything your verifier needs to score the answer. A Resources Server, a Python server you write or reuse, provides the tools and the scoring code that every task shares.

JSONL rows fit many tasks that share one verifier and toolset, such as math, instruction following or tool calling. Tasks that each need their own software environment, such as terminal and coding tasks, fit Harbor Task Folders better.

Data Format

Each JSONL line requires a responses_create_params field following the OpenAI Responses API schema. You don’t need an OpenAI account: local servers such as vLLM use the same format.

{"responses_create_params": {"input": [{"role": "user", "content": "What is 2+2?"}]}}

Additional fields like expected_answer vary by resources server—the component that provides tools and reward signals.

Required Fields

Optional Fields

The agent_ref Field

The agent_ref field maps each row to a specific agent server, which in turn knows its resources server from the YAML config. A training dataset can blend multiple agent servers in a single file—agent_ref tells NeMo Gym which server handles each row.

{
  "responses_create_params": {"input": [{"role": "user", "content": "..."}]},
  "agent_ref": {"type": "responses_api_agents", "name": "math_with_judge_simple_agent"}
}

You don’t create agent_ref manually. The gym dataset collate tool adds it automatically based on your config file. The tool matches the agent type (responses_api_agents) with the agent name from the config.

Example Data

{"responses_create_params": {"input": [{"role": "user", "content": "What is 2+2?"}]}, "expected_answer": "4"}
{"responses_create_params": {"input": [{"role": "user", "content": "What is 3*5?"}]}, "expected_answer": "15"}
{"responses_create_params": {"input": [{"role": "user", "content": "What is 10/2?"}]}, "expected_answer": "5"}

Quick Start

Run this command from the repository root:

gym dataset collate \
    --config responses_api_models/vllm_model/configs/vllm_model_for_training.yaml \
    --resources-server example_multi_step \
    --output-dir data/test \
    --mode example_validation

Success: Finished! message and data/test/example_metrics.json created.

Dataset Types

Configuration

Define datasets in your agent server’s YAML config:

datasets:
  - name: train
    type: train
    jsonl_fpath: resources_servers/workplace_assistant/data/train.jsonl
    # Unified dataset source. `type` selects the backend; the other fields are backend-specific.
    source:
      type: huggingface
      repo_id: nvidia/Nemotron-RL-agent-workplace_assistant
      artifact_fpath: train.jsonl
    license: Apache 2.0

Dataset source

source is the unified way to declare where a dataset is fetched from. type selects the backend; the remaining fields are backend-specific:

# Hugging Face Hub
source:
  type: huggingface
  repo_id: nvidia/Nemotron-RL-agent-workplace_assistant
  artifact_fpath: train.jsonl
# GitLab dataset registry
source:
  type: gitlab
  dataset_name: example_multi_step
  version: 0.0.1
  artifact_fpath: train.jsonl

Valid Licenses

Apache 2.0 · MIT · GNU General Public License v3.0 · Creative Commons Attribution 4.0 International · Creative Commons Attribution-ShareAlike 4.0 International · TBD · NVIDIA Internal Use Only, Do Not Distribute

Workflow

Validation Modes

To prepare training data with auto-download:

gym dataset collate \
    --config responses_api_models/vllm_model/configs/vllm_model_for_training.yaml \
    --resources-server workplace_assistant \
    --output-dir data/workplace_assistant \
    --mode train_preparation \
    --download

Common Errors

Guides

CLI Commands

See CLI Commands for details.

Large Datasets

• Validation streams line-by-line (memory-efficient)
• Single-threaded; >100K samples may take minutes
• Use num_repeats instead of duplicating JSONL lines