README.md

---
dataset_info:
  features:
  - name: text
    dtype: string
  - name: video
    dtype: string
  - name: hard_negative_texts
    list: string
  - name: hard_negative_videos
    list: string
  splits:
  - name: train
    num_bytes: 956287081
    num_examples: 2700
  download_size: 956287081
  dataset_size: 956287081
configs:
- config_name: default
  data_files:
  - split: train
    path: metadata.parquet
---

# Physics Bench Solid With Hard Negatives

Repository: `gowitheflowlab/physics-bench-solid-w-hardnegs`

Source dataset: `gowitheflowlab/physics-bench-solid-train-2700`

This repository follows the path-based layout of `gowitheflowlab/physics-bench-optics-w-hardnegs`.

- rows: 2700
- metadata columns: `text`, `video`, `hard_negative_texts`, `hard_negative_videos`
- positive text: query 1 (`__query_1`) from the source case
- hard negatives per row: 5
- video paths: repository-relative `videos/<family>/<case_id>.mp4`
- list alignment: `hard_negative_texts[i]` and `hard_negative_videos[i]` come from the same case
- candidate pool: only the other 99 cases in the positive case's family

## Reproducibility

- global seed: 42
- traversal: family alphabetical order, then case_id order
- sampling: one `random.Random(42)` stream for all 2700 rows
- selection: `random.sample(sorted(other_99_case_ids), 5)`

## Loading

```python
from pathlib import Path
import pyarrow.parquet as pq
from huggingface_hub import snapshot_download

root = Path(snapshot_download("gowitheflowlab/physics-bench-solid-w-hardnegs", repo_type="dataset"))
rows = pq.read_table(root / "metadata.parquet").to_pylist()
row = rows[0]
positive_video = root / row["video"]
negative_videos = [root / path for path in row["hard_negative_videos"]]
```

The training metadata intentionally contains only the four retrieval columns. Case IDs and query IDs are retained in `source_metadata/sampling_manifest.json` for auditing.