experiment_data / README.md
XueyingJiang's picture
Add files using upload-large-folder tool
2e44f11 verified
|
Raw History Blame Contribute Delete
2.28 kB
# ScanNet v5 axis-white SFT data
This package contains the user-approved ScanNet v5 SFT data used for the Qwen3.5-0.8B training and evaluation runs.
## Contents
- `scannet_v5_axiswhite_edbffd1_stable_train.tar.zst`: stable training split, 8,000 samples and 8,000 PNG images.
- `scannet_v5_axiswhite_edbffd1_stable_eval.tar.zst`: stable evaluation split, 1,000 samples and 1,000 PNG images.
- `scannet_v5_axiswhite_edbffd1_diverse4_train.tar.zst`: diverse4 training split, 8,000 samples and 8,000 PNG images.
- `scannet_v5_axiswhite_edbffd1_diverse4_eval.tar.zst`: diverse4 evaluation split, 1,000 samples and 1,000 PNG images.
- `metadata/`: source manifests.
- `SOURCE_IDENTITY.txt`: source version and accepted exception.
- `SHA256SUMS`: package checksums.
## Download
This public dataset uses manual access approval. Request access on the dataset page first, then authenticate and download:
```bash
hf auth login
hf download XueyingJiang/experiment_data --repo-type dataset --local-dir ./experiment_data
cd ./experiment_data
sha256sum -c SHA256SUMS
```
The download command is resumable. Running it again continues or verifies files already present in the local directory.
## Extraction
Run all commands in the same destination directory:
```bash
tar --use-compress-program=unzstd -xf scannet_v5_axiswhite_edbffd1_stable_train.tar.zst
tar --use-compress-program=unzstd -xf scannet_v5_axiswhite_edbffd1_stable_eval.tar.zst
tar --use-compress-program=unzstd -xf scannet_v5_axiswhite_edbffd1_diverse4_train.tar.zst
tar --use-compress-program=unzstd -xf scannet_v5_axiswhite_edbffd1_diverse4_eval.tar.zst
sha256sum -c SHA256SUMS
```
After extraction, data is arranged under `stable/{train,val}` and `diverse4/{train,val}`.
## Visual convention
For the two local-coordinate vector tasks, the true 0.5 m scale is integrated into one positive X, Y, or Z axis. The segment from the coordinate origin to the 0.5 m tick is white and labeled `0.5 m`.
## Known accepted exception
The user accepted that stable training sample `scannet_v5_stable_train_004274` has a roll reference arc whose projection appears close to a straight line. The full validator was interrupted at this sample, so this package must not be described as having passed that validator without exception.