You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Please share your details to instantly access this sample dataset. Your request is auto-approved.

Log in or Sign Up to review the conditions and access this dataset content.

BLINDSPOT

8 of 9 frontier models score nearly the same on a biology diagram benchmark when the image is completely removed. This is a 10-row sample of the scored responses. Access is auto-approved — fill in your details to view the data.

Full dataset (6,335 responses · 9 models · 4 conditions · 181 items): Nalandadata/BLINDSPOT — requires access request.

The BLINDSPOT benchmark shows an unlabelled scientific diagram to a model and asks it to name every part a leader line points to. The labeled original is the answer key. The control condition removes the image entirely — models only receive the subject name. The finding: most models are reciting from training data, not reading the diagram.


Key finding

grounded contribution = told_score − no_image_score   (percentage points)
recitation share      = no_image_score ÷ told_score   (%)

gpt-4o-mini recitation share: 99% (+0.2 pts grounded, CI contains zero — the image contributes nothing). gpt-5.6-luna recitation share: 9% (+46.2 pts grounded — highest genuine visual contribution).

Columns

Column Type Description
item_id string Public item code (NALANDA-N)
model string Model identifier, e.g. google/gemini-3.6-flash
condition string blind, told, desc, or no-image
f1 float Synonym-aware name-set F1 (0–1)
precision float Precision component
recall float Recall component
n_pred int Number of names the model predicted
n_key int Number of ground-truth names
predicted_names list[str] The model's label list
key_names list[str] Ground-truth anatomical names
subject string Short factual description of the diagram
tier string Item tier (core or extended)
unsolved_core bool True if no baseline model solved this item
recitable bool True if ≥1 model scored F1 ≥ 0.5 with no image
parse_ok bool Whether the model's output parsed correctly

Why no images in the release?

All 181 items come from copyrighted textbooks and a proprietary question bank. Only scores and predicted names are released publicly.


info@nalandadata.ai · nalandadata.ai

Downloads last month
51