README.md

---
pretty_name: SPARC VQA Raw
tags:
  - robotics
  - vision-language
  - spatial-reasoning
  - sparc
---

# SPARC VQA Raw

This repository contains the unfiltered SPARC VQA corpus: 838,211 embedded-image training examples in `train.parquet` (33.20 GB). Each example contains an image, question, answer, task metadata, source identifier, and annotation metadata including `selected_start_score`.

## Ready-to-train version

For the exact processed SPARC subset used by the released Qwen3.5 models, download `train_filtered_t097_mpo700.parquet` from [irl-kit/SPARC-VQA](https://huggingface.co/datasets/irl-kit/SPARC-VQA). It contains 284,909 examples and can be used directly without SPARC postprocessing.

After download, load it as a Parquet dataset:

```python
from datasets import load_dataset

dataset = load_dataset(
  "parquet",
  data_files="train_filtered_t097_mpo700.parquet",
  split="train",
)
```

## Reproduce the release filter

To create the same subset from this unfiltered corpus, download `train.parquet` and run the included `export_sparc_training_subset.py` in the same directory:

```bash
python export_sparc_training_subset.py \
  train.parquet \
  train_filtered_t097_mpo700.parquet \
  --quality-threshold 0.97 \
  --max-per-object 700 \
  --blocked-vacant-location gripper
```

The filter applies `selected_start_score >= 0.97`, sorts retained records by descending score, retains complete trajectory-subtask groups for the first 700 groups per normalized object phrase, and removes vacant-location prompts referring to a gripper.

For a custom raw-data subset, change any of `--quality-threshold`, `--max-per-object`, or `--blocked-vacant-location`, then load the resulting Parquet with the same `load_dataset("parquet", ...)` call above. The output preserves the embedded Hugging Face-compatible image column.

## Data schema

| Field | Description |
| --- | --- |
| `sample_id` | Stable example identifier |
| `image` | Embedded image bytes and optional path |
| `question` | User text prompt |
| `answer` | Supervised assistant answer |
| `target_type`, `task_type` | Spatial target and VQA task labels |
| `source`, `split`, `metadata` | Provenance and generation metadata |

`ours_adaptive_det_soft_snr_sp8.yaml` records the raw SPARC VQA generation configuration. `release_mixture.yaml` records the SPARC and external-dataset mixture used for each released model.

## Citation

```bibtex
@article{blank2026sparc,
  title={SPARC: Reliable Spatial Annotations from Robot Demonstrations at Scale},
  author={Blank, Nils and others},
  journal={arXiv preprint arXiv:2606.13497},
  year={2026}
}
```