Blackbean109/caveman-world-knowledge-150k

---
language:
- en
license: mit
task_categories:
- text-generation
- question-answering
pretty_name: Caveman World Knowledge 150K
size_categories:
- 100K<n<1M
configs:
- config_name: default
data_files:
- split: train
path: train.parquet
- split: validation
path: validation.parquet
---

# Caveman World Knowledge 150K

## Dataset description

Caveman-style instruction dataset with two blended behaviors:

- known world knowledge responses (Wikipedia-like content rewritten in caveman voice)
- unknown-question reactions with mood labels: angry, argue, attack

This dataset is intended for instruction tuning and style conditioning.

## Dataset structure

Each row is a JSON object with fields:

- `id`: unique row id
- `source`: `wikipedia`, `fallback`, or `synthetic`
- `topic`: world topic or `unknown`
- `instruction`: user question/prompt
- `response`: caveman-style answer
- `mood`: `neutral`, `angry`, `argue`, or `attack`
- `knowledge_status`: `known` or `unknown`
- `style`: `caveman`
- `language`: `en`

## Intended use

- style transfer experiments
- robust unknown-question behavior tuning
- synthetic instruction tuning with persona control

## Limitations

- automatically generated paraphrases can contain factual simplifications
- persona language is intentionally ungrammatical
- unknown behavior includes aggressive tone and should be reviewed for deployment suitability

## Generation process

- topic list from `topics_world.txt`
- unknown prompt list from `unknown_questions.txt`
- summaries fetched from Wikipedia when available
- fallback local facts for offline generation
- text rewritten to caveman style via rule-based transforms

## Recommended checks

- sample and manually audit factual quality
- profanity and safety filtering if used in public products
- domain balancing checks (science/history/geography/etc.)