RADAR: Benchmarking Language Models on Imperfect Tabular Data

Link: Paper | Code

The Robust And Data Aware Reasoning (RADAR) benchmark is designed to evaluate the ability of language models to demonstrate data-awareness—that is, to recognize, reason over, and appropriately handle complex data artifacts such as:

• Missing data
• Bad values
• Outliers
• Inconsistent formatting
• Inconsistent multi-column logic

The full dataset includes 53 tasks grounded in real-world data tables and varies across data artifact types and table dimensions (by token count and number of columns). In total, RADAR provides 2,980 unique query-table task instances. We also include two subsets of the data: (1) radar-sizes (RADAR-S) to focus evaluation on table sizes and (2) radar-tasks (RADAR-T) to focus evaluation across all tasks.

📊 Dataset Statistics

Dataset Split	Tasks	Instances	Tokens (K)	Cols
RADAR	53	2,980	[2,4,8,16]	[5,10,20]
RADAR-T	53	313	8	10
RADAR-S	10	720	[2,4,8,16]	[5,10,20]

🔭 Dataset Structure

Each task instance comprises of the follwowing data:

• task_id: a unique id for each source table and query
• query: the query to ask over the data table
• answer: ground truth answer to the query
• artifact_type: the artifact type introduced to the data table for this task
• artiact_scope: does reasoning over the data artifacts involve only a single column, naively or independetly over multiple columns, or jointly or connected over multiple columns
• query_cols: the columns invovled in the query
• artifact_reasoning_cols: the columns invovled in reasoning over the artifacts
• table: the data table for this task (a dictionary with keys "headers" and "rows" to represent the table column names and rows)
• num_rows: number of rows in the tbale
• num_cols: number of columns in the table
• recovered_tables_transform_spec: The right answer is caluclated over the recovered data table(s). We convert the data table in table to the recovered data table(s) using this specification indicating which rows to drop and which cells to overwrite.
• base_data_num_tokens: The number of tokens in the data table (before introducing any data artifact perturbations). This may be slightly different after introducing perturbations.
• base_data_token_bucket: The token bucket in which this task belongs to (one of 2000, 4000, 8000, and 16000)
• perturbation_note: Any note about the data artifact perturbation that is introduced.

💻 Loading the Data

Using Hugging Face

from datasets import load_dataset
radar_all = load_dataset("kenqgu/radar", "radar")["test"]
radar_s = load_dataset("kenqgu/radar", "radar-sizes")["test"]
radar_t = load_dataset("kenqgu/radar", "radar-tasks")["test"]

Using included RADAR code to load into more usable pydantic objects (need to install radar first).

from radar.data import load_task_instances_hf

# load the full dataset
tasks, task_summary_df = load_task_instances_hf(split="full")
tasks_s, _ = load_task_instances_hf(split="sizes")
tasks_t, _ = load_task_instances_hf(split="tasks")

# view the table as a pandas dataframe
tasks[0].table_df.head()

📖 Citation

If you use RADAR in your research, please cite our paper:

@article{gu2025radar,
  title={RADAR: Benchmarking Language Models on Imperfect Tabular Data}, 
  author={Ken Gu and Zhihan Zhang and Kate Lin and Yuwei Zhang and Akshay Paruchuri and Hong Yu and Mehran Kazemi and Kumar Ayush and A. Ali Heydari and Maxwell A. Xu and Girish Narayanswamy and Yun Liu and Ming-Zher Poh and Yuzhe Yang and Mark Malhotra and Shwetak Patel and Hamid Palangi and Xuhai Xu and Daniel McDuff and Tim Althoff and Xin Liu},
  year={2025},
  eprint={2506.08249},
  archivePrefix={arXiv},
  primaryClass={cs.DB},
  url={https://arxiv.org/abs/2506.08249}, 
}