"```python\nfrom modal_training_gym.common.dataset import HuggingFaceDataset\n```\n\nDataset backed by a HuggingFace `datasets` repo.\n\nSubclass and set `hf_repo` plus column mappings. When\n`input_column` and `output_column` are set, `prepare()` wraps\neach row into a prompt-only chat message list plus a separate label\nfield: `{\"messages\": [{\"role\": \"user\", ...}], <label_key>: ...}`.\nA leading `{\"role\": \"system\", ...}` message is included when\n`system_prompt` is set. No assistant turn is emitted — the target\nfrom `output_column` is stored under `label_key`.\n\n**Inherits from:** `DatasetConfig`\n\n## Fields\n\n| Field | Type | Default | Description |\n|-------|------|---------|-------------|\n| `dataset_id` | `str` | `\"\"` | |\n| `input_key` | `str` | `\"\"` | |\n| `label_key` | `str` | `\"label\"` | |\n| `apply_chat_template` | `bool` | `True` | |\n| `always_prepare` | `bool` | `False` | |\n| `writes_eval_paths` | `bool` | `True` | |\n| `hf_repo` | `str` | `\"\"` | |\n| `hf_split` | `str` | `\"train\"` | |\n| `hf_config` | `str \\| None` | `None` | |\n| `output_format` | `str` | `\"parquet\"` | |\n| `input_column` | `str` | `\"\"` | |\n| `output_column` | `str` | `\"\"` | |\n| `system_prompt` | `str` | `\"\"` | |\n| `prompt_template` | `str` | `\"{input}\"` | |\n| `n_rows` | `int` | `0` | |\n\n## Methods\n\n### `load(self, split: \"Literal['all', 'train', 'eval']\" = 'all') -> 'Any'`\n\nLoad raw examples, optionally filtered by split.\n\n### `prepare(self, path: 'str', eval_paths: 'dict[str, str] | None' = None) -> 'None'`\n\nMaterialize training data to `path` (and eval splits to `eval_paths`).\n\n### `to_pandas(self, *, formatted: 'bool' = False)`\n\n### `validate_prepared(self, path: 'str') -> 'None'`\n\nSniff what `prepare()` wrote and confirm the columns the framework will index.\n\n## Related Tutorials\n\n- [DAPO on math with Qwen3-4B](/tutorials/rl/005_dapo/)\n- [Sweep hyperparameters across runs with TrainingGroup](/tutorials/rl/007_param_sweep/)\n\n**Source:** [`modal_training_gym/common/dataset.py`](https://github.com/modal-projects/training-gym/blob/main/modal_training_gym/common/dataset.py)\n"

HuggingFaceDataset

Dataset backed by a HuggingFace datasets repo.

from modal_training_gym.common.dataset import HuggingFaceDataset

Dataset backed by a HuggingFace datasets repo.

Subclass and set hf_repo plus column mappings. When input_column and output_column are set, prepare() wraps each row into a prompt-only chat message list plus a separate label field: {"messages": [{"role": "user", ...}], <label_key>: ...}. A leading {"role": "system", ...} message is included when system_prompt is set. No assistant turn is emitted — the target from output_column is stored under label_key.

Inherits from: DatasetConfig

Fields

Field	Type	Default
`dataset_id`	`str`	`""`
`input_key`	`str`	`""`
`label_key`	`str`	`"label"`
`apply_chat_template`	`bool`	`True`
`always_prepare`	`bool`	`False`
`writes_eval_paths`	`bool`	`True`
`hf_repo`	`str`	`""`
`hf_split`	`str`	`"train"`
`hf_config`	`str \| None`	`None`
`output_format`	`str`	`"parquet"`
`input_column`	`str`	`""`
`output_column`	`str`	`""`
`system_prompt`	`str`	`""`
`prompt_template`	`str`	`"{input}"`
`n_rows`	`int`	`0`

Methods

`load(self, split: "Literal['all', 'train', 'eval']" = 'all') -> 'Any'`

Load raw examples, optionally filtered by split.

`prepare(self, path: 'str', eval_paths: 'dict[str, str] | None' = None) -> 'None'`

Materialize training data to path (and eval splits to eval_paths).

`to_pandas(self, *, formatted: 'bool' = False)`

`validate_prepared(self, path: 'str') -> 'None'`

Sniff what prepare() wrote and confirm the columns the framework will index.

Source: modal_training_gym/common/dataset.py