apat1n/UltraData-Math-boxed

---
language:
- en
license: apache-2.0
task_categories:
- text-generation
tags:
- math
- grpo
- rl
- boxed
size_categories:
- 100K<n<1M
configs:
- config_name: default
data_files: data/train-*
default: true
- config_name: normalized
data_files: normalized/train-*
---

# UltraData-Math-boxed

High-confidence subset of [UltraData/UltraData-Math](https://huggingface.co/datasets/UltraData/UltraData-Math) containing only problems with `\boxed{}` answers.

**Original dataset**: [UltraData/UltraData-Math](https://huggingface.co/datasets/UltraData/UltraData-Math) (53M samples across 2 configs)

## Variants

| Config | Samples | Description |
|--------|---------|-------------|
| **default** | 390,082 | Raw `\boxed{}` extractions, minimal cleanup |
| **normalized** | 217,049 | Filtered + validated: no units, no variables, no multi-value, sympy LaTeX verified |

## Why boxed-only?

UltraData-Math is synthetic textbook data where only ~2-7% of solutions contain `\boxed{}` final answers. The remaining ~93% require regex extraction from solution text (e.g. grabbing the last `= something`), which produces ~1-5% wrong ground truth labels. For RL training (GRPO), incorrect ground truth creates false learning signals.

This dataset keeps only the high-confidence `\boxed{}` extractions for clean reward signal.

## Normalized Variant Filters

1. **Answer cleaning**: Strip LaTeX delimiters, trailing junk, degree symbols
2. **Answer validation**: Must contain digits, <=30 chars, no variables, no multi-value, no units, no prose, no broken LaTeX
3. **LaTeX verification** via sympy `parse_latex`
4. **Problem filtering**: No multi-part, no proofs, no sketching, no rewrite tasks, length 30-2000 chars

## Stats

| Source Config | Total Rows | With boxed | Rate |
|---|---|---|---|
| L3-Textbook-Exercise-Synthetic | 26M | ~411k | 1.6% |
| L3-QA-Synthetic | 27M | ~1.87M | 6.9% |
| **After dedup** | **53M** | **390k** | **0.7%** |
| **After normalization** | - | **217k** | - |

## Dataset Format

| Column | Type | Description |
|--------|------|-------------|
| problem | str | Math problem text |
| answer | str | Answer extracted from `\boxed{}` in the solution |

## Usage

```python
from datasets import load_dataset

# Raw boxed extractions (390k)
ds = load_dataset("apat1n/UltraData-Math-boxed")

# Normalized + filtered (217k)
ds = load_dataset("apat1n/UltraData-Math-boxed", "normalized")
```

## Credits

Based on [UltraData/UltraData-Math](https://huggingface.co/datasets/UltraData/UltraData-Math).