Add stratified-subset README

README.md

@@ -0,0 +1,138 @@
+---
+dataset_info:
+  features:
+  - name: file_name
+    dtype: string
+  - name: source_file
+    dtype: string
+  - name: question
+    dtype: string
+  - name: question_type
+    dtype: string
+  - name: question_id
+    dtype: int32
+  - name: answer
+    dtype: string
+  - name: answer_choices
+    list: string
+  - name: correct_choice_idx
+    dtype: int32
+  - name: image
+    dtype: image
+  - name: video
+    dtype: video
+  - name: media_type
+    dtype: string
+  splits:
+  - name: test
+    num_examples: 10343
+configs:
+- config_name: default
+  data_files:
+  - split: test
+    path: data/test-*
+license: mit
+task_categories:
+- visual-question-answering
+language:
+- en
+size_categories:
+- 10K<n<100K
+tags:
+- engineering
+- simulation
+- stratified-subset
+---
+
+# OpenSeeSimE-Structural-Small
+
+A **stratified 10% subset** of [`cmudrc/OpenSeeSimE-Structural`](https://huggingface.co/datasets/cmudrc/OpenSeeSimE-Structural), intended for quicker iteration, smoke tests, and baseline training on reduced data.
+
+## Subset Provenance
+
+- **Parent dataset**: [`cmudrc/OpenSeeSimE-Structural`](https://huggingface.co/datasets/cmudrc/OpenSeeSimE-Structural) (102,678 rows total)
+- **Rows in this subset**: **10,343** (10.07% of parent)
+- **Parquet shards**: 4   |   **Storage**: ~15.60 GB
+- **Sampling**: per-stratum shuffle with `numpy.random.default_rng(42)`, then take `ceil(n * fraction)` from each stratum. Any non-empty stratum contributes at least 1 row.
+- **Strata**: `(source_file, question_type, media_type, question_id)` — all four jointly.
+- **Nesting**: `Mini ⊂ Small ⊂ parent`. The same shuffled prefix is taken for every fraction, so the 1% subset is a literal subset of the 10% subset.
+
+## Composition
+
+### By `source_file`
+
+| source_file     |   rows |   pct |
+|:----------------|-------:|------:|
+| Beams           |   2088 | 20.19 |
+| Pressure Vessel |   2074 | 20.05 |
+| Dog Bone        |   2061 | 19.93 |
+| Hip Implant     |   2060 | 19.92 |
+| Wall Bracket    |   2060 | 19.92 |
+
+### By `media_type`
+
+| media_type   |   rows |
+|:-------------|-------:|
+| image        |   5192 |
+| video        |   5151 |
+
+### By `(source_file, question_type)`
+
+| source_file     |   Binary |   Multiple Choice |   Spatial |   Total |
+|:----------------|---------:|------------------:|----------:|--------:|
+| Beams           |      627 |              1045 |       416 |    2088 |
+| Dog Bone        |      619 |              1030 |       412 |    2061 |
+| Hip Implant     |      618 |              1030 |       412 |    2060 |
+| Pressure Vessel |      622 |              1040 |       412 |    2074 |
+| Wall Bracket    |      618 |              1030 |       412 |    2060 |
+
+## Deviations from Parent
+
+- **`Beam` → `Beams` relabel**: 20 rows in the parent dataset were mislabeled `source_file='Beam'` (singular) when they should have been `'Beams'` (plural). In this subset, those rows have been merged into `'Beams'`. The 5 `source_file` classes in this subset are: `Beams`, `Dog Bone`, `Hip Implant`, `Pressure Vessel`, `Wall Bracket`.
+
+## Feature Schema
+
+Identical to the parent dataset. See [`cmudrc/OpenSeeSimE-Structural`](https://huggingface.co/datasets/cmudrc/OpenSeeSimE-Structural) for full documentation of simulation generation, ground-truth extraction, preprocessing, limitations, and intended use.
+
+```python
+{
+    'file_name':          str,        # Unique identifier
+    'source_file':        str,        # Base simulation model
+    'question':           str,        # Question text
+    'question_type':      str,        # 'Binary', 'Multiple Choice', 'Spatial'
+    'question_id':        int,        # Question identifier (1-20)
+    'answer':             str,        # Ground truth answer
+    'answer_choices':     list[str],  # Options
+    'correct_choice_idx': int,        # Index of correct answer
+    'image':              Image,      # PIL Image (1920x1440) or null for video rows
+    'video':              Video,      # Video bytes or null for image rows
+    'media_type':         str,        # 'image' or 'video'
+}
+```
+
+## Intended Use
+
+- Fast smoke-testing of VLM evaluation pipelines before running the full benchmark
+- Baseline training (e.g. CNN baselines) on a reduced footprint
+- Iteration on metric/aggregation code without 100k+ row overhead
+
+For rigorous benchmarking, use the parent dataset.
+
+## License
+
+MIT — same as parent. Free for academic and commercial use with attribution.
+
+## Citation
+
+```bibtex
+@article{ezemba2024opensesime,
+  title={OpenSeeSimE: A Large-Scale Benchmark to Assess Vision-Language Model Question Answering Capabilities in Engineering Simulations},
+  author={Ezemba, Jessica and Pohl, Jason and Tucker, Conrad and McComb, Christopher},
+  year={2025}
+}
+```
+
+## Contact
+
+**Jessica Ezemba** — jezemba@andrew.cmu.edu  
+Department of Mechanical Engineering, Carnegie Mellon University