XLGym

XLGym contains natural-language spreadsheet tasks paired with Office.js reference programs, input and reference-output workbooks, and recorded candidate outcomes.

Splits

Split	Tasks	Input filenames	Distinct input byte contents	Target APIs
train	20,482	1,722	814	1,287
bench	200	126	122	82

Artifact	Train	Bench
Chart	3,520	25
ConditionalFormat	3,512	25
Filter	1,686	25
Formula	3,517	25
Hyperlink	191	25
PivotTable	3,500	25
Sort	1,068	25
Table	3,488	25

Load the records

After publication:

from datasets import load_dataset

dataset = load_dataset("prose-ms/xlgym")
train = dataset["train"]
bench = dataset["bench"]

The same package can be loaded before publication without contacting the Hub:

dataset = load_dataset(
    "parquet",
    data_files={
        "train": r"data\train.parquet",
        "bench": r"data\bench.parquet",
    },
)

The data files contain metadata and paths, not embedded XLSX bytes.

Download and extract workbooks

Three ZIP archives avoid publishing tens of thousands of individual files:

Archive	Contents
workbooks/inputs.zip	1,848 referenced input workbook filenames across both splits
workbooks/train_outputs.zip	20,482 reference-output workbooks
workbooks/bench_outputs.zip	200 reference-output workbooks and 1,000 candidate outputs

Archive members are relative to a common extraction root. Their paths match input_workbook, gold_wb, and solution_wb_dir in each record.

from pathlib import Path
from zipfile import ZipFile
from huggingface_hub import snapshot_download

local = Path(snapshot_download(
    "prose-ms/xlgym",
    repo_type="dataset",
    allow_patterns=["data/*", "workbooks/*", "*.json", "*.md", "LICENSE", "SHA256SUMS"],
))
books = Path("xlgym_workbooks")
for name in ("inputs.zip", "train_outputs.zip", "bench_outputs.zip"):
    with ZipFile(local / "workbooks" / name) as archive:
        archive.extractall(books)

row = bench[0]
input_path = books / row["input_workbook"]
reference_path = books / row["gold_wb"]

For benchmark-only use, omit train_outputs.zip. XLSX files are already compressed; the ZIP archives use stored entries. workbook_manifest.json contains SHA-256 checksums and byte sizes for every member. SHA256SUMS covers the package files.

Training candidates retain code, execution-success flags, attempt counts and stdout, but their output workbooks are intentionally not bundled.

Record fields

See SCHEMA.md for the full contract. Core fields include:

• query: task given to a solver.
• input_workbook: input workbook path after archive extraction.
• query_solution_code: historical reference Office.js program.
• expected_answer: author-proposed expected result.
• gold_wb, gold_stdout: saved reference workbook and execution output.
• artifact, api_name: generation labels, not proof that a particular API is necessary or actually used.
• candidates, pass_rate: historical candidate attempts and grading outcomes.

When evaluating solvers, provide only the query, input workbook and the context your protocol permits. Do not expose reference code, expected answers, reference outputs, candidate solutions or author rationale as solver input. The benchmark includes these fields for evaluation and inspection, not training.

Generation and grading

Tasks were mined from API/workbook pairs over EnronFUSE workbooks. An author proposed a query, expected answer and reference program. Consistency checking and candidate-solution testing could trigger revisions. The historical pipeline supplied API information to candidate solvers; this release should not be described as an API-blind mining experiment.

The retained label records at most two passes among five candidates: pass_rate is 0.0, 0.2 or 0.4. Candidate execution and model-based grading are part of this measurement. Exact historical protocol revisions are not fully pinned per record.

These scores are not intrinsic task-difficulty measurements or independent correctness certificates. Known historical limitations include incomplete workbook evidence, output-policy conflicts, reference-only grading constraints, and an extraction issue that could omit cell differences after context building. Later pilot corrections did not retroactively regenerate or reverify these rows.

This export preserves the recorded supervision and labels rather than silently relabeling them. Reference and candidate code should be inspected and executed only in an appropriately isolated environment. Running Office.js requires a compatible Excel environment; the dataset itself does not include a portable Excel runtime or a standalone grading service.

Provenance and limitations

• Source package: historical xlgymv2_22k; original bench200 is named bench here.
• No task filtering, resampling, or validation holdout was applied during this packaging step.
• Added fields provide stable identifiers, portable archive paths and workbook hashes. Source text, programs and candidate outcomes are preserved.
• Only referenced workbooks are included. Intermediate query-evolution traces, raw logs, credentials, unrelated source trees and pilot runs are not bundled.
• Original workbooks and generated text may contain names, addresses or other source-corpus content; they should not be assumed anonymized.
• The intended use is spreadsheet-agent research. Review supervision quality for applications requiring reliable correctness.

License

This dataset is licensed under the Community Data License Agreement - Permissive, Version 2.0 (CDLA-Permissive-2.0). See LICENSE for the complete terms. When sharing the data, include the license text.