---
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: 598672486
    num_examples: 2700
  download_size: 598672486
  dataset_size: 598672486
configs:
- config_name: default
  data_files:
  - split: train
    path: metadata.parquet
---

# Physics Bench Optics Train

Repository: `gowitheflowlab/physics-bench-optics-train`

Training split with hard negatives for optics simulation video retrieval.

- rows: 2700
- metadata columns: `text`, `video`, `hard_negative_texts`, `hard_negative_videos`
- hard negatives per row: 5, drawn from the other 99 cases of the same family
- 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
- `text` and `hard_negative_texts` use the natural-language query style (`v9_optics_dynamics_aligned_stable_phrasing`),
  matching the `parsed_text` column of `gowitheflowlab/physics-bench-optics-eval-2700`
- case parameters are disjoint from the evaluation split, so no evaluation case is reachable here

## Reproducibility

- sampling: global seed 42, family alphabetical order then case_id order, one
  `random.Random(seed)` stream, `random.sample(sorted(other_99_case_ids), 5)`
- assignment digest: `b4e45b4e1e5c3faa`
- text: generated from the training case metadata; see
  `source_metadata/build_hardnegs_v2_metadata.py` and `source_metadata/v2_manifest.json`
- validation report: `quality/validation.json`

## Loading

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

root = Path(snapshot_download("gowitheflowlab/physics-bench-optics-train", 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"]]
```