Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -5,4 +5,21 @@ configs:
|
|
| 5 |
data_files:
|
| 6 |
- split: train
|
| 7 |
path: data.zip
|
|
|
|
|
|
|
| 8 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 5 |
data_files:
|
| 6 |
- split: train
|
| 7 |
path: data.zip
|
| 8 |
+
source_datasets:
|
| 9 |
+
- Pokerme/view2space-train
|
| 10 |
---
|
| 11 |
+
|
| 12 |
+
# view2space
|
| 13 |
+
|
| 14 |
+
VIEW2SPACE's native training MCQs, grouped by ordered source images.
|
| 15 |
+
|
| 16 |
+
Source: Pokerme/view2space-train at 1af36aa39a18143db6599a049da6de657e18215c.
|
| 17 |
+
Only explicit multiple-choice questions are included; counting and detection are
|
| 18 |
+
excluded. Original PNG bytes, image order, question IDs, input boxes and reasoning
|
| 19 |
+
are preserved. Reasoning is metadata, never question input. Related QAs share one
|
| 20 |
+
image group to avoid repeating image bytes for every question. The source has only
|
| 21 |
+
a training split; no evaluation split is fabricated. The source README states CC BY 4.0.
|
| 22 |
+
Duplicate distractors are deduplicated with gold indices remapped; ambiguous duplicate
|
| 23 |
+
gold options are excluded and recorded in excluded-questions.jsonl.
|
| 24 |
+
|
| 25 |
+
Original data: [Pokerme/view2space-train](https://huggingface.co/datasets/Pokerme/view2space-train). Repackaged as native ImageFolder for [tasksource](https://github.com/sileod/tasksource) by [scripts/upload_repackaged.py](https://github.com/sileod/tasksource/blob/main/scripts/upload_repackaged.py).
|