sumuks/explanation-prefs-2.5k

---
pretty_name: Explanation Prefs 2.5k
language:
- en
license: mit
task_categories:
- text-generation
tags:
- synthetic
- dpo
- preference-optimization
- explanations
- analogy
- education
size_categories:
- 1K<n<10K
configs:
- config_name: default
data_files:
- split: train
path: train-00000-of-00001.parquet
- split: test
path: test-00000-of-00001.parquet
---

# Dataset Card for explanation-prefs-2.5k

## Dataset Summary

Explanation Prefs 2.5k is a small synthetic preference dataset for explanation-style alignment experiments.
Each example contains a short educational question, a `chosen` explanation that incorporates an analogy naturally, and a `rejected` explanation that answers the same question without an analogy.
The dataset is intended for testing data pipelines, preference training code, and small-scale experiments rather than for benchmarking factual knowledge.

## Dataset Structure

- Train rows: 2500
- Test rows: 500
- Total rows: 3000
- Generation model: `gpt-4.1-mini`
- Global generation seed: `7`

Each row contains these fields:

- `domain`: Source concept domain such as `biology` or `programming`.
- `template`: Prompt template family.
- `concept_a`: Primary concept in the question.
- `concept_b`: Optional secondary concept for comparison prompts.
- `prompt`: Natural-language question shown to the model.
- `chosen`: Preferred explanation that incorporates an analogy naturally.
- `rejected`: Less preferred explanation that avoids analogies.
- `seed`: Stable per-row generation seed.
- `source_model`: The model used to generate the pair.

## Generation Process

Prompt texts are generated locally from curated concept lists and simple educational templates such as:

- `What is X?`
- `What is X, and what is Y?`
- `What is the difference between X and Y?`
- `When would someone use X instead of Y?`

For each prompt, the script makes two separate OpenAI calls.
The `chosen` call uses the base question plus an extra instruction to explain with an analogy.
The `rejected` call uses the same base question plus an instruction to answer plainly without analogies.
The `chosen` answer should incorporate an analogy naturally into the explanation.
The `rejected` answer must explain the same topic without analogies, metaphors, or similes.
Rows are validated for schema, length, and the required analogy distinction before they are written.

## Intended Uses

- Testing DPO or other preference-optimization pipelines.
- Small synthetic experiments about explanation style.
- Smoke tests for dataset loading, filtering, and prompt formatting.

## Limitations

- The dataset is fully synthetic and inherits the biases and mistakes of the generator model.
- The preference signal is narrow: analogy versus no analogy.
- It should not be treated as a benchmark of explanation quality or factual coverage.

## Loading Example

```python
from datasets import load_dataset

dataset = load_dataset(
"parquet",
data_files={"train": "train-00000-of-00001.parquet", "test": "test-00000-of-00001.parquet"},
)
print(dataset)
```