🤗 Dataset To Data List

The loader already gives you a rows Data List - so why does this node exist? Because rows is only the untransformed dataset. The moment you chain Filter, Map Column, Shuffle, or Take onto the loader's dataset output, you have a refined dataset with no Data List attached. 🤗 Dataset To Data List is the bridge: it materializes any dataset the pack produces - loader output or a whole transform chain - into the same Data List shape rows had, so you can shape first and convert last.

How it works

Three inputs, one output:

• dataset - whatever HUGGINGFACE_DATASET you've built up.
• column - empty means one item per row dict; set it and each item is that column's value instead.
• limit - caps how many rows get materialized (-1 = all).

The output rows is a ComfyUI Data List: the companion Basic data handling pack treats it as one item per row, and it distinguishes two ways of using it. Whole-list nodes (Data List → length, get item, first) receive the entire list in one call. Any other Basic node - a DICT → get, a STRING op - runs once per row, and its output is a new Data List of results. That per-row mapping is the idiomatic way to transform every record in one go, and it's exactly why this converter exists.

The canonical pattern

Load (streaming) → 🤗 Dataset Filter (label == 1) → 🤗 Dataset To Data List (limit=50) → Data List → length

Keep only the positive reviews, count them - and because the loader is streaming and the converter's limit is 50, almost nothing actually downloads. This shape-first, materialize-last ordering is the pack's whole design thesis, and To Data List is where a chain of 🤗 nodes finally hands off to the generic ComfyUI data world.

To Data List vs. To LIST

Both converters take the same inputs and both materialize; the difference is the output shape and what it's good for. To Data List gives per-row processing (a node runs once per row) - right for mapping every row through a transform. To LIST gives one flat Python list - right for whole-list operations like length, get item, or feeding a single column of values into string nodes. Pick the converter that matches what your downstream node expects; switching later is a one-node swap.

Gotchas

Materialization is the moment laziness ends. limit = -1 on a non-streaming dataset has already paid for the full download at the loader, so the converter just builds the list; on a streaming dataset, limit is your real brake - set it or put a Take first, or you may pull far more of the stream than you wanted. Row values are plain Python dicts keyed by column names, with numpy scalars converted for you; image/audio objects pass through unchanged. And spell column exactly - a typo surfaces as a missing-key error on row one.

Installing it

Part of StableLlama/ComfyUI-huggingface_dataset. ComfyUI-Manager → search "Hugging Face dataset", or:

cd ComfyUI/custom_nodes
git clone https://github.com/StableLlama/ComfyUI-huggingface_dataset
pip install -r requirements.txt

Restart ComfyUI. The Basic data handling pack is a separate install - grab it too, since these converters are built around its Data List semantics.