Pravallika6/hit_dataset_grouped_v2

---
configs:
- config_name: default
data_files:
- split: train
path: train.parquet
- split: validation
path: val.parquet
- split: test
path: test.parquet
---

# HiT Grouped Dataset v2

Extended grouped dataset with sibling and grandparent columns for full loss ablation.

## Columns
- `child` — question or trigger observation
- `parent_1 parent_2 parent_3` — true parents (padded to 3)
- `negative_1 negative_2 negative_3 negative_4 negative_5 negative_6 negative_7 negative_8` — sampled negatives from same-depth pool
- `sibling_1 sibling_2` — another child of parent_i (for sibling sector loss)
- `grandparent_1 grandparent_2` — 2-hop ancestors (for transitive cone loss)

## Usage
```python
HierarchyTransformerLoss(
n_parents=3, n_negatives=8,
n_siblings=2, n_grandparents=2
)
```