Datasets:
File size: 1,974 Bytes
c7a1cd4 1716e03 c7a1cd4 4b4183f c7a1cd4 1716e03 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 | ---
pretty_name: V-Zero 5K
license: other
task_categories:
- visual-question-answering
size_categories:
- 1K<n<10K
tags:
- multimodal
- reinforcement-learning
- on-policy-distillation
- v-zero
---
# V-Zero 5K
This repository contains 5,013 multimodal training examples for V-Zero.
The training image columns are `images`, `teacher_images`, `teacher_neg_images`,
and the ablation column `teacher_random_images`.
The published Parquet intentionally omits `data_source`, `ability`,
`reward_model`, and `extra_info`.
All published image paths are relative to the repository root. Run
`materialize_paths.py` after downloading to create a parquet containing
machine-local absolute paths for the V-Zero training launcher.
## Download
Official endpoint:
```bash
hf download hao05/v-zero-5k --repo-type dataset --local-dir ./v-zero-5k
```
Mainland China download mirror:
```bash
HF_ENDPOINT=https://hf-mirror.com \
hf download hao05/v-zero-5k --repo-type dataset --local-dir ./v-zero-5k
```
Extract and verify the image shards:
```bash
python ./v-zero-5k/extract_image_shards.py --root ./v-zero-5k
```
The 16 TAR shards are uncompressed because the contained JPEG/PNG files are
already compressed. Keeping TAR uncompressed avoids recompression overhead and
supports deterministic SHA-256 verification.
Materialize local paths:
```bash
python ./v-zero-5k/materialize_paths.py \
--input ./v-zero-5k/data/train.parquet \
--output ./v-zero-5k/data/train.local.parquet \
--root ./v-zero-5k
```
## Source composition
| Source | Examples |
| --- | ---: |
| `chartqa` | 314 |
| `docvqa` | 437 |
| `gqa_global` | 14 |
| `gqa_relation` | 320 |
| `perception` | 3,297 |
| `reasonseg` | 93 |
| `tallyqa` | 277 |
| `textvqa` | 261 |
## License and responsible release
This is a mixed-source dataset. Users are responsible for checking the
redistribution terms, attribution requirements, generated-annotation terms,
privacy constraints, and source citations before reuse.
|