NAPEval Spreadsheet Trajectories

Benchmark trajectories for predictive auto-completion in spreadsheets. Each record is one spreadsheet-building session represented as an ordered sequence of symbolic cell operations. Given a prefix of a trajectory, the task is to predict the next operation(s) the user will perform.

This is the evaluation data for the next_action_pred_eval framework. See the project page and paper.

Dataset summary

• 52 trajectories (single test split — this is an evaluation benchmark).
• 11907 total operations; per-trajectory length ranges 35–821.
• 23 operation types: ALIGN_HORIZONTAL, ALIGN_VERTICAL, AUTOFILL, BORDER_ALL, BORDER_BOTTOM, BORDER_INSIDE_VERTICAL, BORDER_LEFT, BORDER_OUTSIDE, BORDER_RIGHT, BORDER_TOP, FILL_COLOR, FONT_BOLD, FONT_COLOR, FONT_ITALIC, FONT_NAME, FONT_SIZE, FONT_UNDERLINE, INPUT, MERGE, NUMBER_FORMAT, PASTE_FROM, TEXT_ORIENTATION, WRAP_TEXT.

Fields

Field	Type	Description
name	string	Trajectory id (stable, anonymised).
operations	list[string]	Ordered editing history; each entry is one symbolic operation.
num_operations	int	Convenience length of operations.

The symbolic DSL

Every spreadsheet edit is a pipe-delimited string:

OPERATION_TYPE | Sheet!Range | value

Examples:

Symbolic string	Meaning
INPUT | Sheet1!A1 | "Title"	Set A1 to a string value
INPUT | Sheet1!A1:B2 | [[1, 2], [3, 4]]	Bulk-input a 2×2 block
FONT_BOLD | Sheet1!A5 | True	Bold A5
MERGE | Sheet1!A1:G1 | true	Merge a range
BORDER_TOP | Sheet1!A8:G8 | Thin, Continuous, #000000	Top border on a row

The canonical parsing/serialisation logic lives in the framework's core/operations/ module.

Usage

from datasets import load_dataset

ds = load_dataset("Tej-a55/napeval", split="test")
print(ds[0]["name"], ds[0]["num_operations"])
prefix = ds[0]["operations"][:10]   # history fed to a solver

Supplementary raw artifacts

Each trajectory also ships per-trajectory raw artifacts under raw/<id>/:

File	Description
spreadsheet.xlsx	The finished workbook (rebuilt from the operations).
sheet_image.png	A rendered image of the finished sheet.
predictable_state.json	Annotations: per-cell predictable properties + coverage stats.
operations.txt	Plain-text mirror of operations (one op per line).

These are supplementary — not required to run the benchmark. They live outside the parquet dataset (so they do not appear in the dataset viewer). Download them with:

from huggingface_hub import snapshot_download

snapshot_download("Tej-a55/napeval", repo_type="dataset", allow_patterns="raw/*")

License

Creative Commons Attribution 4.0 International (CC BY 4.0).