README.md

---
pretty_name: Nordet Social Cognition Benchmark
language:
  - fr
  - en
task_categories:
  - question-answering
  - visual-question-answering
tags:
  - benchmark
  - social-cognition
  - quebec
  - multimodal
size_categories:
  - n<1K
---

# Nordet Social Cognition Benchmark

Nordet evaluates social cognition in culturally and historically situated Quebec contexts.
It contains 381 challenge and control examples across four tasks:

- literal vs. intent;
- historical and social inference;
- communication adaptation;
- cultural understanding.

The dataset combines text, transcribed spoken language, and 57 embedded images. Questions are
provided in French and English. Every row includes a reference answer, source citation, and a
reproducible LLM-as-a-judge prompt template.

## Loading

```python
from datasets import load_dataset

dataset = load_dataset("Pythonner/nordet", split="train")
```

For multimodal examples, `image` is decoded as a PIL image. Text-only examples contain `None`.
The `messages` column follows the Hugging Face conversational vision format; an image content
marker identifies where the corresponding `image` belongs in the user message.

## Variants

The `variant` column distinguishes `challenge` from `control`. Controls preserve the general
question and evaluation shape while reducing the culturally situated social inference required.

## Evaluation

Replace `{response_text}` in `llm_as_a_judge_prompt_template` with the evaluated model response:

```python
judge_prompt = row["llm_as_a_judge_prompt_template"].replace(
    "{response_text}", model_response
)
```

The template requests a structured result for every criterion. Reference answers are included in
the criteria where the original Kaggle benchmark used them for judge calibration.

## Sources and rights

Questions were authored for Nordet. Reference answers are grounded in the cited source material.
Images retain their original source URLs and should be used according to the rights identified by
their respective source repositories, including BAnQ and Wikimedia Commons.

See the final project writeup:
<https://www.kaggle.com/competitions/kaggle-measuring-agi/writeups/nordet-social-cognition-benchmark>