# LaTeX Formula Bucket Indices (Train/Test Split)

## Overview

This dataset contains bucket indices for the LaTeX formulas dataset, split into train and test sets.

**Source Dataset:** OleehyO/latex-formulas
**Total Samples:** 552,340
**Split Ratio:** 80% train / 20% test
**Random Seed:** 42

## Split Statistics

- **Train Set:** 441,869 samples (80.0%)
- **Test Set:** 110,471 samples (20.0%)

## Bucket Distribution

### Train Set
- 2line: 73,316 samples
- long: 31,543 samples
- long_2line: 23,364 samples
- long_matrix: 652 samples
- long_single_line: 7,527 samples
- matrix: 6,126 samples
- medium: 264,458 samples
- medium_2line: 47,114 samples
- medium_matrix: 4,437 samples
- medium_single_line: 212,907 samples
- short: 145,868 samples
- short_2line: 2,838 samples
- short_matrix: 1,037 samples
- short_single_line: 141,993 samples
- single_line: 362,427 samples

### Test Set
- 2line: 18,330 samples
- long: 7,886 samples
- long_2line: 5,841 samples
- long_matrix: 163 samples
- long_single_line: 1,882 samples
- matrix: 1,533 samples
- medium: 66,116 samples
- medium_2line: 11,779 samples
- medium_matrix: 1,110 samples
- medium_single_line: 53,227 samples
- short: 36,469 samples
- short_2line: 710 samples
- short_matrix: 260 samples
- short_single_line: 35,499 samples
- single_line: 90,608 samples

## Files

- `indices/indices_*_train.json` - Training set indices per bucket
- `indices/indices_*_test.json` - Test set indices per bucket
- `indices/split_mapping.json` - Complete train/test split mapping

## Usage

```python
import json
from huggingface_hub import hf_hub_download

# Download specific bucket
filepath = hf_hub_download(
    repo_id="wulanbhai/subset",
    filename="indices/indices_short_train.json",
    repo_type="dataset"
)

with open(filepath, 'r') as f:
    indices = json.load(f)

print(f"Loaded {len(indices):,} training indices for 'short' bucket")
```

## Bucket Definitions

- **Length buckets:**
  - short: ≤50 tokens
  - medium: 51-150 tokens
  - long: >150 tokens

- **Structure buckets:**
  - single_line: Single-line formulas
  - 2line: Multi-line formulas (using `\\`)
  - matrix: Matrix structures

- **Combined buckets:** Intersection of length × structure (e.g., `short_single_line`)

## Important Notes

⚠️ **For Fair Comparison:**
- Use **train set** for model training/fine-tuning
- Use **test set** for evaluation ONLY
- Never train on test set!
- Never test on train set!

✅ **Stratified Split:**
- Each bucket maintains 80%/20% ratio
- Distribution preserved across all buckets

## Citation

If you use these indices, please cite the original dataset:
```
@dataset{latex_formulas,
  author = {OleehyO},
  title = {LaTeX Formulas Dataset},
  year = {2024},
  url = {https://huggingface.co/datasets/OleehyO/latex-formulas}
}
```

## License

Indices only (no original data redistributed).