RAVEC-data-backup / README.md
Jack04810's picture
Add RAVEC/PAPO reproducibility metadata and source manifest
5a8bd5a verified
|
Raw History Blame Contribute Delete
1.51 kB
---
pretty_name: RAVEC/PAPO public metadata backup
tags:
- multimodal
- reproducibility
- metadata
---
# RAVEC data backup
This public dataset repository stores the reproducible metadata boundary for
RAVEC/PAPO. It does not mirror image bytes: the official PAPO datasets already
provide the image-bearing files, and their dataset cards should be followed
for the applicable terms.
## Sources
- Training: [PAPOGalaxy/PAPO_ViRL39K_train](https://huggingface.co/datasets/PAPOGalaxy/PAPO_ViRL39K_train), 38,870 rows.
- Legacy MMK12 monitoring: [PAPOGalaxy/PAPO_MMK12_test](https://huggingface.co/datasets/PAPOGalaxy/PAPO_MMK12_test).
- Evaluation: [PAPO-Galaxy/PAPO_eval](https://huggingface.co/datasets/PAPO-Galaxy/PAPO_eval).
The upstream cards did not expose an explicit license field when this backup
was prepared, so this repository keeps metadata and source pointers only.
## Files
`metadata/full_train.jsonl` and `metadata/clean_eval_2k.jsonl` are path-neutral
copies of the local records. Absolute paths are replaced with logical image
IDs; sample IDs, answers, source rows, image hashes and RAVEC annotation fields
are preserved. `MANIFEST.json` records counts, sizes and SHA256 values.
Run `bash scripts/prepare_from_upstream.sh /path/to/data` to download the
official image-bearing sources directly. The script clears proxy environment
variables and does not upload or modify local code.
No credentials, checkpoints, optimizer/EMA/RNG state, private logs or absolute
local paths are included.