Prism MathBlend Curated

This release combines normalized mathematics instruction records from two direct upstream datasets. Of the released records, 45% come from databricks/databricks-dolly-15k, 45% come from openai/gsm8k, and the remaining 10% are newly written schema examples.

The percentages above describe the records actually included in this release. Normalization only changed field names and whitespace.

For quality measurement, the maintainers used Qwen/Qwen2.5-7B-Instruct to score answer clarity. That model did not generate, rewrite, or contribute any released record.

Project utilities are hosted in the Prism-LicenseDesk repository, and companion releases are listed under toolathlonEval.