---
license: mit
pretty_name: TopBench
task_categories:
  - tabular-classification
  - tabular-regression
  - question-answering
language:
  - en
tags:
  - tabular
  - benchmark
  - predictive-reasoning
  - llm
---

# TopBench Dataset

TopBench is a benchmark for predictive reasoning over tabular data. Each example asks a model to infer an unobserved outcome, decision, treatment effect, or ranked/filtering result from historical tables and a natural-language query.

## Layout

```text
single_point_prediction/
decision_making/
treatment_effect_analysis/
ranking_and_filtering/
```

Each task directory contains dataset folders with `history.csv`, query JSON files, and metadata. The `ranking_and_filtering` task also includes `current.csv` and expects structured CSV outputs.

Use the code release at https://github.com/LAMDA-Tabular/TopBench for validation, inference, evaluation, and the predict-only baseline.