---
license: mit
task_categories:
  - robotics
tags:
  - VLA
  - vision-language-action
  - pruning
  - calibration
---

# VLADrop Calibration Data

Calibration sets for the layer-importance profiling in
**Drop-Then-Recovery (DTR): How Redundant Are Vision-Language-Action Models?**
([paper](https://arxiv.org/abs/2606.27755) · [code](https://github.com/s1ghhh/VLADrop) ·
[checkpoints](https://huggingface.co/collections/s1ghhh/vladrop-drop-then-recovery-dtr-checkpoints-6a509dd598cf54ae53060204)).

Each file holds the exact 512 samples (64 batches × 8) a profiling run consumes — not the full
dataset. Loading these reproduces the paper's GateProbe / baseline-metric block rankings exactly.

| File | Setting | Seed |
|---|---|---|
| `pi05_libero_dropped_calib_64x8_seed42.pt` | pi0.5 × LIBERO | 42 |
| `pi05_libero_plus_calib_64x8_seed42.pt` | pi0.5 × LIBERO-Plus | 42 |
| `openvla_libero_calib_64x8_seed9999.pt` | OpenVLA-OFT × LIBERO | 9999 |
| `openvla_libero_plus_calib_64x8_seed9999.pt` | OpenVLA-OFT × LIBERO-Plus | 9999 |

pi0.5 calibration is stored in bfloat16 (pi0.5's runtime precision) and is regenerated
deterministically by `profiling/pi0.5/dump_calib.py` in the code repo.

## Usage

```bash
huggingface-cli download s1ghhh/VLADrop_Calibration_Data --repo-type dataset \
    --local-dir profiling/calibration_data
```

See `profiling/README.md` in the [code repo](https://github.com/s1ghhh/VLADrop) for the full
reproduction commands.