LTEM-STB: Synthetic Table Benchmark

LTEM-STB is the Synthetic Table Benchmark (STB) of LTEM: LLM-Friendly Table-Enhanced Markdown for Table Understanding and Generation (EMNLP 2026), a table question-answering benchmark. It controls table structure so that formats can be compared level by level.

• Paper: https://arxiv.org/abs/2610.29188
• Project page: https://hugman.github.io/ltem/
  • Code and LTEM specification
• EurekaAI: https://eurekax.co/

Overview

• Tables: 1,089 synthetic tables
• Questions: 4,356, four per table (LOOKUP, COMPARISON, COUNTING, AGGREGATION)
• Formats: CSV, Markdown, MultiMarkdown, HTML, LaTeX, LTEM

Complexity levels

Level	Name	Tables	Structure
L1	Atomic grid	96	Rows and columns only
L2	Headered	334	A header row
L3	Spanning	342	Merged cells
L4	Multi-level header	164	Headers nested under merged parent headers
L5	Block-partitioned	87	Section rows that split a table into blocks
L6	Composite	66	Several sub-tables with different columns in one unit

L1–L4 build on each other; L5 and L6 add section stacking or table composition on top of simpler tables.

Data

from datasets import load_dataset
qa = load_dataset("eunbinkim777/ltem-stb", "qa", split="test")
tables = load_dataset("eunbinkim777/ltem-stb", "tables", split="test")

• qa: one row per question, with its answer and the cells that hold it (targets)
• tables: one row per table, with its cell structure and the table in all six formats
• data/: the same data as one file per table (index.csv, tables/, formats/<format>/, qa/)