add-dataset

name	add-dataset
description	Register or inspect a Hugging Face dataset for Marin pipelines.

Skill: Dataset Schema Inspection and Registration

Overview

Inspect a Hugging Face dataset schema with Marin's schema inspection tool, then add a dataset module under experiments/datasets/ so the dataset can be tokenized and consumed in Marin pipelines. Each leaf module exposes <name>_dataset() (one corpus) or <name>_datasets() -> dict[str, ...] (a keyed family) built with the lazy data builders in marin.experiment.data.

• Single-corpus datasets: add a small leaf module (e.g. experiments/datasets/svg.py) exposing a <name>_dataset().
• Multipart/complex datasets: create a dedicated file (e.g. experiments/datasets/nemotron.py) exposing <name>_datasets().
• Datasets with HF-exposed subsets/splits: pattern-match on nemotron.py, keying one handle per subset in the returned dict.

Prerequisites

• Prefer repo-managed dependencies over ad hoc pip install.
• For repeated work in a checked-out Marin repo, install the synced environment with uv sync --all-packages, then run uv run lib/marin/tools/get_hf_dataset_schema.py.
• For one-off schema inspection without a provisioned environment, use ephemeral deps: uv run --with datasets --with pyyaml lib/marin/tools/get_hf_dataset_schema.py ...
• Ensure access to the dataset (Hugging Face Hub ID, local path, or other supported format).

Usage

Command line:

uv run lib/marin/tools/get_hf_dataset_schema.py <dataset_name> [options]

Python import:

from marin.tools.get_hf_dataset_schema import get_schema
schema = get_schema(dataset_name="wikitext", config_name="wikitext-103-v1")

Rules

1. Config handling: always check whether a dataset requires a config first. If required, the tool returns {"error": "Config name is required.", "available_configs": [...]} — select an appropriate config from the list and retry with --config_name.
2. Text field selection: prioritize a field named exactly text; fall back to fields containing text; consider string-type fields if no obvious text field exists. Examine sample_row to verify field contents.
3. Error handling: handle missing config, dataset not found, and remote code execution required. Retry with appropriate parameters (e.g. --trust_remote_code).
4. Performance: the tool streams to avoid full downloads; expect quick responses. sample_row may be empty for some datasets.

Output Format

The tool returns a JSON object:

{
  "splits": ["train", "validation", ...],
  "text_field_candidates": ["text", "content", ...],
  "features": {
    "text": "string",
    "label": "int64",
    ...
  },
  "sample_row": {
    "text": "Example content...",
    ...
  }
}

Example: dataset requiring a config

$ uv run lib/marin/tools/get_hf_dataset_schema.py wikitext
{
  "error": "Config name is required.",
  "available_configs": ["wikitext-103-raw-v1", "wikitext-103-v1", ...]
}

$ uv run lib/marin/tools/get_hf_dataset_schema.py wikitext --config_name wikitext-103-v1
{
  "splits": ["train", "validation", "test"],
  "text_field_candidates": ["text"],
  "features": {"text": "string"},
  "sample_row": {"text": "Article content..."}
}

For datasets needing remote code, add --trust_remote_code.

Next Steps

Once the schema is inspected and the dataset is registered, cargo-cult existing dataset configs for tokenization:

• Apply transformations (e.g. field mapping).
• Estimate token counts and file sizes.
• Find similar dataset configurations in Marin's existing experiments.
• Copy and adapt tokenization configs from similar datasets.
• Run ablations or trials.

See Also

• lib/marin/tools/get_hf_dataset_schema.py
• Hugging Face datasets documentation