Caveman-Style World Knowledge Dataset for Instruction Tuning

Caveman World Knowledge 150K is an instruction dataset containing approximately 150,000 entries for tuning language models. It was created by author Blackbean109 and was last updated in April 2026. The dataset blends factual world knowledge responses with reactions to unknown questions.

Use Cases

• Instruction tuning models to generate responses in a specific 'caveman' voice using the `source` and `topic` fields for content conditioning.
• Training classifiers to detect and react to unknown questions based on the provided mood labels like 'angry', 'argue', or 'attack'.
• Analyzing the distribution and style of rewritten factual content across different `source` types such as 'wikipedia', 'fallback', or 'synthetic'.
• Creating synthetic dialogue agents with distinct personality traits by conditioning on the blended behavioral data and `topic` categories.

Strengths

• Approximately 150,000 instruction-response pairs based on the dataset name.
• Provides structured JSON data with distinct fields like `id`, `source`, and `topic` for filtering.

Limitations

• The exact row count, column details, and sample data are not provided in the input.
• Potential class imbalance or bias in the mood labels ('angry', 'argue', 'attack') is unknown.

Provenance

Source

Content is derived from sources including Wikipedia, with fallback and synthetic entries.

Collection Method

World knowledge was rewritten into a 'caveman' voice and blended with synthetic unknown-question reactions.

Freshness

Last updated on 2026-04-07.