# Attribution and modifications

OpenAI: [GSM8K / grade-school-math](https://github.com/openai/grade-school-math), revision `3101c7d5072418e28b9008a6636bde82a006892c`, original `grade_school_math/data/train.jsonl`, MIT. The complete upstream MIT copyright and permission notice is retained in `licenses/GSM8K-MIT.txt`. Background: [Training Verifiers to Solve Math Word Problems](https://arxiv.org/abs/2110.14168).

Glaive AI: [glaive-function-calling-v2](https://huggingface.co/datasets/glaiveai/glaive-function-calling-v2/tree/e7f4b6456019f5d8bcb991ef0dd67d8ff23221ac), revision `e7f4b6456019f5d8bcb991ef0dd67d8ff23221ac`, Apache-2.0 as declared by the source card. The source card is retained in `provenance/glaive-source-card.md`.

Modifications by thenukegun10x: selected and screened source tasks; added a numeric-answer instruction to math prompts; normalized conversation/tool representations; retained source clarification context where present; replaced target continuations with independently generated Liquid AI BF16 teacher reasoning plus verified answers/calls; made grouped train/validation splits; rendered and tokenized the accepted data; created this Parquet release and documentation. The generated continuations are not original OpenAI/Glaive reference explanations.

Teacher and formatting template: Liquid AI, [LFM2.5-24B-A2B](https://huggingface.co/LiquidAI/LFM2.5-24B-A2B/tree/1769516ac588c4c558089c94f27d96b0dc67b183), revision `1769516ac588c4c558089c94f27d96b0dc67b183`. The template is copied unmodified, with its upstream license retained.