| --- |
| license: cc |
| task_categories: |
| - text-generation |
| tags: |
| - mathematica |
| - textgen |
| - synthetic |
| pretty_name: HALF |
| size_categories: |
| - 10M<n<100M |
| --- |
| ### HALF Mathematic Database |
| ### grow... |
|
|
| | **Feature** | **HALF** | **GSM8K** | **MATH** | |
| | -------------------- | ---------------------------------------------- | -------------------------------------- | -------------------------------------- | |
| | **Primary Focus** | Pre-training / Synthetic Symbolic Pre-training | Grade school math word problems | Competition-level mathematics | |
| | **Dataset Size** | **~53.9 Million examples** (4 shards) | ~8.5K examples (7.5K train / 1K test) | 12.5K examples (7.5K train / 5K test) | |
| | **Format** | Single-line LaTeX equations (`A \implies B`) | Natural language word problems | LaTeX & step-by-step reasoning proofs | |
| | **Natural Language** | **None** (pure LaTeX math expressions) | High (English narrative word problems) | High (English proofs + LaTeX notation) | |
| | **Data Generation** | Programmatically generated / Synthetic | Human-written / Human-verified | Human-compiled (AMC, AIME, HMMT) | |
| | **Target Task** | Symbolic identity training, syntax alignment | Multi-step arithmetic reasoning | Deep conceptual problem-solving | |
