mahmad1882/blind-spots

Overview

Blind Spots is a benchmark for examining how language models respond when they encounter uncertain, ambiguous, culturally situated, or insufficiently specified questions. It investigates whether models distinguish between what they know, what they can reasonably infer, and what they do not know.

The benchmark explores how model behavior changes across prompt framings and language conditions, including English, Urdu, and Roman Urdu. It is designed to investigate failure modes that can be missed by conventional evaluations focused primarily on whether a final answer matches a reference answer.

The central research questions are:

1. Knowledge boundaries and framing: When a model encounters a question it cannot reliably answer, does it acknowledge the limits of its knowledge, or does its response depend on how the question is framed and expressed?
2. Domain and language variation: How does a model's ability to express uncertainty appropriately vary across knowledge domains (e.g., culturally niche versus general knowledge) and languages, including English and lower-resource language settings such as Urdu and Roman Urdu?

Motivation

This dataset investigates these behaviors through variations in prompt framing and language, as well as multi-turn interactions that test whether a model maintains or reverses a response when challenged. To a more limited extent, it also examines how enabling or disabling a thinking mode affects these behaviors.

The goal is to characterize the conditions under which models provide reliable answers, appropriately acknowledge uncertainty, or produce unsupported claims, and to identify and to identify how these behaviors vary across domains, languages, and interaction conditions.

Dataset contents

The repository provides two CSV files.

1. data/blind_spots_dataset.csv

The benchmark input dataset, containing containing 1,904 samples spanning 75 underlying item identifiers, 10 framing or interaction conditions, and 4 language or language-style conditions. Not every item appears in every combination of framing and language.

Column	Description
sample_id	Sample identifier.
pool	Evaluation pool or category.
framing	Prompt-framing condition.
language	Language or language-style condition.
prompt	Input prompt presented to the model.
gold	Reference answer or expected target.
options	Candidate options or relevant answer constraints, where applicable.

2. data/blind_spots_evaluation.csv

The evaluation dataset, containing 2,276 rows. A limited subset of samples is prompted again with thinking 'on' hence the discrepancy between the number of rows in the two datasets.

In addition to the benchmark inputs, it contains model responses and evaluation annotations, including:

• response: The model's response to the prompt.
• LABEL: The assigned evaluation label.
• FABRICATED: A binary label for presence or absence of fabricated knowledge.
• notes: Extra details, where relevant.
• Additional fields recording sample identifiers, evaluation modes, and relevant evaluation conditions.

Refer to the CSV headers for the complete schema.

Experimental setup

The benchmark was developed to study a small open-weight language model under controlled prompting conditions.

• Model: Qwen3.5-4B
• Model repository: Qwen/Qwen3.5-4B
• Evaluation inputs: Prompts from data/blind_spots_dataset.csv
• Evaluation outputs: Model responses and evaluation annotations in data/blind_spots_evaluation.csv

For reproducibility and detailed experimental configurations, check the README under the collection this dataset belongs to.

Loading the dataset

Install the required library:

pip install -U datasets

Load either dataset configuration:

from datasets import load_dataset

repo_id = "mahmad1882/blind-spots"

# Load the benchmark prompts and reference answers
prompts = load_dataset(repo_id, "prompts", split="train")

# Load model responses and evaluation annotations
evaluation = load_dataset(repo_id, "evaluation", split="train")

print(prompts)
print(evaluation)

The dataset is hosted at mahmad1882/blind-spots.

Alternatively, download the CSV files directly:

• Benchmark input dataset
• Evaluation dataset