ARC-Style Reasoning Prompts and Instruction Pairs

ARC Mega is a large-scale mixture of Abstraction and Reasoning Corpus (ARC) style prompts and non-ARC instruction/answer pairs. The dataset, created by mindware, originated from the jack-arc training runs and was last updated in November 2025. It is split into ARC-only and non-ARC partitions while preserving the 'MindsAI' prompt/response CSV format.

Use Cases

• Fine-tune language models on ARC-style prompts to improve performance on abstract reasoning tasks.
• Train models on the non-ARC instruction/answer pairs to enhance general instruction-following capabilities.
• Use the canonical JSON representation of ARC tasks for structured evaluation of model reasoning.
• Filter and analyze the dataset partitions (ARC-only vs. non-ARC) to study different facets of model intelligence.
• Benchmark model performance on the 'MindsAI' prompt/response format for consistency with specific training runs.

Strengths

• Dataset is described as a 'large-scale mixture' of ARC and non-ARC data.
• Provides a structured split between ARC-only and non-ARC partitions for focused analysis.
• Includes a minified JSON representation specifically for canonical ARC tasks.

Limitations

• Specific row count, column names, and total dataset size are unknown.
• Rows containing synthetic dataset-helper prompts, system/user/GPT prefixes, or malformed ARC reconstructions were filtered out, indicating potential data quality issues in the raw source.
• The description is incomplete, referencing a full description on an external page.

Provenance

Source

mindware, originating from jack-arc training runs.

Collection Method

Mixture of ARC-style prompts and non-ARC instruction/answer pairs; some rows filtered for quality.

Freshness

Last updated on 2025-11-05.