metadata
license: cc-by-4.0
configs:
- config_name: posterior_sampled_100_estimation
data_files:
- split: test
path: posterior_sampled_100_estimation/test-*
- config_name: posterior_sampled_100_most_freq
data_files:
- split: test
path: posterior_sampled_100_most_freq/test-*
- config_name: posterior_sampled_100_second_freq
data_files:
- split: test
path: posterior_sampled_100_second_freq/test-*
- config_name: posterior_sampled_50_estimation
data_files:
- split: test
path: posterior_sampled_50_estimation/test-*
- config_name: posterior_sampled_50_most_freq
data_files:
- split: test
path: posterior_sampled_50_most_freq/test-*
- config_name: posterior_sampled_50_second_freq
data_files:
- split: test
path: posterior_sampled_50_second_freq/test-*
- config_name: prior_estimation
data_files:
- split: test
path: prior_estimation/test-*
- config_name: prior_most_freq
data_files:
- split: test
path: prior_most_freq/test-*
- config_name: prior_second_freq
data_files:
- split: test
path: prior_second_freq/test-*
dataset_info:
- config_name: posterior_sampled_100_estimation
features:
- name: qid
dtype: string
- name: task
dtype: string
- name: domain
dtype: string
- name: distribution_type
dtype: string
- name: condition
dtype: string
- name: source
dtype: string
- name: answer
dtype: string
- name: ref_dist
dtype: string
- name: question
dtype: string
- name: meta_data
dtype: string
- name: comments
dtype: string
splits:
- name: test
num_bytes: 8231382
num_examples: 324
download_size: 1932613
dataset_size: 8231382
- config_name: posterior_sampled_100_most_freq
features:
- name: qid
dtype: string
- name: task
dtype: string
- name: domain
dtype: string
- name: distribution_type
dtype: string
- name: condition
dtype: string
- name: source
dtype: string
- name: answer
dtype: string
- name: ref_dist
dtype: string
- name: question
dtype: string
- name: meta_data
dtype: string
- name: comments
dtype: string
splits:
- name: test
num_bytes: 8072412
num_examples: 320
download_size: 1905166
dataset_size: 8072412
- config_name: posterior_sampled_100_second_freq
features:
- name: qid
dtype: string
- name: task
dtype: string
- name: domain
dtype: string
- name: distribution_type
dtype: string
- name: condition
dtype: string
- name: source
dtype: string
- name: answer
dtype: string
- name: ref_dist
dtype: string
- name: question
dtype: string
- name: meta_data
dtype: string
- name: comments
dtype: string
splits:
- name: test
num_bytes: 8005961
num_examples: 317
download_size: 1926447
dataset_size: 8005961
- config_name: posterior_sampled_50_estimation
features:
- name: qid
dtype: string
- name: task
dtype: string
- name: domain
dtype: string
- name: distribution_type
dtype: string
- name: condition
dtype: string
- name: source
dtype: string
- name: answer
dtype: string
- name: ref_dist
dtype: string
- name: question
dtype: string
- name: meta_data
dtype: string
- name: comments
dtype: string
splits:
- name: test
num_bytes: 5278674
num_examples: 324
download_size: 1158909
dataset_size: 5278674
- config_name: posterior_sampled_50_most_freq
features:
- name: qid
dtype: string
- name: task
dtype: string
- name: domain
dtype: string
- name: distribution_type
dtype: string
- name: condition
dtype: string
- name: source
dtype: string
- name: answer
dtype: string
- name: ref_dist
dtype: string
- name: question
dtype: string
- name: meta_data
dtype: string
- name: comments
dtype: string
splits:
- name: test
num_bytes: 5046088
num_examples: 315
download_size: 1154778
dataset_size: 5046088
- config_name: posterior_sampled_50_second_freq
features:
- name: qid
dtype: string
- name: task
dtype: string
- name: domain
dtype: string
- name: distribution_type
dtype: string
- name: condition
dtype: string
- name: source
dtype: string
- name: answer
dtype: string
- name: ref_dist
dtype: string
- name: question
dtype: string
- name: meta_data
dtype: string
- name: comments
dtype: string
splits:
- name: test
num_bytes: 5004501
num_examples: 312
download_size: 1148719
dataset_size: 5004501
- config_name: prior_estimation
features:
- name: qid
dtype: string
- name: task
dtype: string
- name: domain
dtype: string
- name: distribution_type
dtype: string
- name: condition
dtype: string
- name: source
dtype: string
- name: answer
dtype: string
- name: ref_dist
dtype: string
- name: question
dtype: string
- name: meta_data
dtype: string
splits:
- name: test
num_bytes: 1618866
num_examples: 324
download_size: 276509
dataset_size: 1618866
- config_name: prior_most_freq
features:
- name: qid
dtype: string
- name: task
dtype: string
- name: domain
dtype: string
- name: distribution_type
dtype: string
- name: condition
dtype: string
- name: source
dtype: string
- name: answer
dtype: string
- name: ref_dist
dtype: string
- name: question
dtype: string
- name: meta_data
dtype: string
splits:
- name: test
num_bytes: 1510402
num_examples: 316
download_size: 265880
dataset_size: 1510402
- config_name: prior_second_freq
features:
- name: qid
dtype: string
- name: task
dtype: string
- name: domain
dtype: string
- name: distribution_type
dtype: string
- name: condition
dtype: string
- name: source
dtype: string
- name: answer
dtype: string
- name: ref_dist
dtype: string
- name: question
dtype: string
- name: meta_data
dtype: string
splits:
- name: test
num_bytes: 1508040
num_examples: 312
download_size: 263521
dataset_size: 1508040
📌 Overview
Text2DistBench is a reading comprehension benchmark for evaluating whether large language models can infer distributional knowledge from natural language evidence. Given metadata and a set of user comments about an entity (e.g., a movie or song), models must infer statistics such as: stance/topic distribution, and most/second- frequent labels.
⚙️ Configurations
The dataset is constructed from movie and music entities released between 2026-01-01 and 2026-04-01. It includes two types of configurations:
🟦 Posterior Configurations
Include both metadata and user comments as evidence.
These configurations differ by sample size (50 or 100 comments) and task type (estimation, most_freq, second_freq).
Format: posterior_sampled_<comment_num>_<task>
🟨 Prior Configurations
Include only metadata (no comments).
Format: prior_<task>
⚙️ Dataset Format
Each benchmark instance corresponds to a distributional reading comprehension question.
{
"qid": < question id >,
"qtype": < distribution type >,
"answer": < answer (mode label or distribution depending on task) >,
"ref_dist": < Ground-truth distribution over labels >,
"question": < Full prompt shown to model (instruction + evidence + query) >,
"source": "< entity name >",
"meta_data": "< text evidence >",
"comments": "< text evidence (if posterior)>",
"condition": "< conditioning variable for P(s|t), P(t|s) >",
}
⚙️ Load the Dataset
from datasets import load_dataset
ds = load_dataset("frett/Text2DistBench", "posterior_sampled_50_estimation", split="test")
ds = load_dataset("frett/Text2DistBench", "prior_most_freq", split="test")