ChronoBias

Dataset Description

ChronoBias is a benchmark for evaluating time-conditional group bias in the time-sensitive knowledge of Large Language Models (LLMs).

• Paper: ChronoBias: A Benchmark for Evaluating Time-conditional Group Bias in the Time-sensitive Knowledge of Large Language Models
• Venue: Findings of EMNLP 2025
• Authors: Kyungmin Kim, Youngbin Choi, Hyounghun Kim, Dongwoo Kim, Sangdon Park

Overview

The benchmark focuses on two types of bias:

• Parametric Knowledge Bias: Bias in group-specific knowledge encoded in model parameters
• Time-sensitivity Bias: Bias in the degree of time-sensitivity across groups

The dataset covers football league standings from multiple international leagues, enabling controlled evaluation of LLM performance across groups (leagues) and time periods.

---

Dataset Statistics

Field	Value
Total examples	13,502
Group	6 (EPL, K-League1, Saudi Pro League, Ligue 1 Algeria, Serie A Brazil, Primera Liga Spain)
Time-sensitivity	3 (never_changing, slow_changing, moderate_to_fast_changing)

---

Data Fields

Field	Type	Description
id	string	Unique identifier for the example
question_id	string	Identifier for the question
league_name	string	Name of the football league
season_name	string	Season year
round	string	Match round number
question	string	The question asked to the LLM
answer	list	Ground-truth answer (list of [team, points] pairs)
context	string	HTML table with league standings as context (for RAG)
current_date	string	The date at which the question is posed (YYYY-MM-DD)
start_date	string	Start date of the round (YYYY/MM/DD)
end_date	string	End date of the round (YYYY/MM/DD)
end_month	string	End month of the round (YYYY/MM)
qtype	string	Question subtype (e.g., top_1, bottom_3)
qtype_agg	string	Aggregated question type (top or bottom)
question_type	string	Question category
time_type	string	How fast this type of information changes over time

---

Usage

from datasets import load_dataset

dataset = load_dataset("tml-lab/ChronoBias")
data = dataset["train"]

# Filter by league
epl_data = data.filter(lambda x: x["league_name"] == "EPL")

# Filter by question type
top1_data = data.filter(lambda x: x["qtype"] == "top_1")

# Use context field for RAG evaluation
example = data[0]
print(example["question"])   # Question to the LLM
print(example["context"])    # HTML table with league standings
print(example["answer"])     # Ground-truth answer

---

Citation

@inproceedings{kim-etal-2025-chronobias,
    title = "{C}hrono{B}ias: A Benchmark for Evaluating Time-conditional Group Bias in the Time-sensitive Knowledge of Large Language Models",
    author = "Kim, Kyungmin  and
      Choi, Youngbin  and
      Kim, Hyounghun  and
      Kim, Dongwoo  and
      Park, Sangdon",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2025",
    month = nov,
    year = "2025",
    address = "Suzhou, China",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.findings-emnlp.405/",
    doi = "10.18653/v1/2025.findings-emnlp.405",
    pages = "7658--7693",
}

---

License

This dataset is released under the MIT License.