TableSuite-1K

TableSuite-1K benchmarks predictive and language-grounded tabular intelligence over 1,000 OpenML-referenced datasets.

Task	Input	Evaluation
Prediction	ICL rows or a partially labelled serialized table	classification and regression
Table grounding	a provided table plus a lookup/comprehension question	exact displayed-table facts
Table QA	a provided subtable plus a typed question	programmatic operations

This repository contains metadata and value-free task plans. It contains no OpenML source values, labels, rendered questions, gold answers, model outputs, embeddings, or checkpoints. OpenML remains the source-table distributor.

Configurations

Configuration	Purpose
datasets	source identity, schema, target, split, and provenance
table_prediction_tasks	prediction eligibility and primary metrics
prediction_episodes	frozen prediction query anchors
table_grounding	official provided-table grounding plans
table_question_answering	official programmatic QA plans

Quickstart

Install the matching package and materialize only the OpenML sources you need:

pip install \
  'tablesuite[local,hf,openml] @ git+https://github.com/Sichao-Li/TableSuite-1K.git@v2.1.0'

tablesuite fetch-openml \
  --reference Lester1996/TableSuite-1K \
  --revision v2.1.0 \
  --output openml-parquet \
  --dataset-id openml_45069 \
  --accept-source-terms

from tablesuite import TableSuite

suite = TableSuite.open(
    "Lester1996/TableSuite-1K",
    source="openml-parquet",
    revision="v2.1.0",
)

task = suite.official(
    "table_grounding",
    split="dataset_test",
    dataset_ids=("openml_45069",),
)
example = task[0]
score = task.score(example.id, model(example.prompt))

Prediction uses the same frozen queries across interfaces and a deterministic nested support schedule:

prediction = suite.prediction(
    "icl",
    support=(0.0, 0.1, 0.3, 0.5, 0.7, 0.9, 1.0),
    dataset_ids=("openml_45069",),
)

For model context limits, PredictionDataset.fit_context selects the largest support prefix that fits the actual tokenized prompt and emits an auditable coverage report. It never silently truncates.

Evaluation Contracts

• Prediction is inference-only; no per-dataset parameter updates are allowed.
• Query targets are always private; only selected support labels are visible.
• Grounding and QA operate only on the displayed table slice.
• Wording and gold are generated deterministically from local source data.
• Dataset transfer uses duplicate-aware dedup_cluster_id partitions.
• Results must report response coverage and, for prediction, requested/realized support plus context coverage.

The v2.1 semantic tasks use literal source headers. This release does not claim curated cross-dataset ontology equivalence.

Source Terms

Each referenced OpenML dataset retains its upstream terms. openml_license_claim is provenance metadata, not a license granted by TableSuite-1K. The repository-level other designation reflects heterogeneous source terms.

Code and full protocol documentation: https://github.com/Sichao-Li/TableSuite-1K