integrations.nemo_gym.dataset

integrations.nemo_gym.dataset

Dataset loading for NeMo Gym JSONL files.

Converts NeMo Gym JSONL format into HuggingFace Datasets compatible with TRL’s GRPOTrainer. Supports multi-environment routing via: 1. Per-dataset server_name (all rows in a file go to one server) 2. Per-row agent_ref.name (each row specifies its own server)

Functions

Name Description
load_nemo_gym_datasets Load and merge NeMo Gym JSONL datasets with multi-environment support.

load_nemo_gym_datasets

integrations.nemo_gym.dataset.load_nemo_gym_datasets(gym_dir, dataset_configs)

Load and merge NeMo Gym JSONL datasets with multi-environment support.

Each dataset config should have

  • path: JSONL file path (absolute, or relative to gym_dir)
  • server_name: Default NeMo Gym server for this dataset. Can be overridden per-row if the JSONL has an “agent_ref” field.
  • max_samples (optional): Max number of samples to use from this dataset

Per-row routing: If a JSONL row has an “agent_ref”: {“name”: “…”} field, that takes precedence over the dataset-level server_name. This allows mixing environments within a single dataset file (matching TRL’s pattern).

The output dataset has columns

  • prompt: list[dict] chat format
  • resources_server_ref: dict with {“name”: server_name}
  • verify_extra: dict with original JSONL data for verify requests

Parameters

Name Type Description Default
gym_dir str Path to the NeMo Gym directory. required
dataset_configs list[dict] List of dataset configuration dicts. required

Returns

Name Type Description
Dataset A HuggingFace Dataset ready for GRPOTrainer.