---
library_name: transformers
pipeline_tag: image-text-to-text
tags:
- robotics
- vision-language
- spatial-reasoning
- sparc
datasets:
- irl-kit/SPARC-VQA
---

# SPARC-Qwen3.5-4B

Qwen3.5-4B fully fine-tuned for embodied spatial reasoning using VQA data generated from SPARC annotations.

## Training data

The training mixture contains SPARC-generated VQA data from `ours_adaptive_det_soft_snr_sp8`, FSD, RoboPoint, and LLaVA-OneVision2. SPARC samples use an annotation-quality threshold of 0.97, are sorted by score, and are capped at 700 samples per object. The paper reports 1,159,047 training pairs for this mixture.

| Release | [SPARC VQA (filtered)](https://huggingface.co/datasets/irl-kit/SPARC-VQA) | FSD | RoboPoint | LLaVA-OneVision2 | EO-1.5M |
| --- | --- | --- | --- | --- | --- |
| Qwen3.5-4B | Yes | Yes | Yes | Yes | No |
| Qwen3.5-0.8B-VTFT | Yes | Yes | Yes | Yes | No |
| Qwen3.5-9B-EO | Yes | Yes | Yes | Yes | Yes |

The SPARC training data is the [filtered release subset](https://huggingface.co/datasets/irl-kit/SPARC-VQA), which needs no further SPARC filtering. Its filtering script and release mixture manifest are in [irl-kit/SPARC-VQA-Raw](https://huggingface.co/datasets/irl-kit/SPARC-VQA-Raw). FSD, RoboPoint, LLaVA-OneVision2, and EO-1.5M remain their respective upstream datasets.

## Prompting

Prompt formatting is important for these models. Use the bundled `chat_template.jinja` through `processor.apply_chat_template(..., add_generation_prompt=True)`, with one `user` turn containing the image(s) followed by the text question. Disable thinking/reasoning mode to match evaluation.

For a single point, append exactly:

```text
Output the point coordinates in JSON format like [{"point_2d": [x, y], "label": "target"}]. Use integer coordinates between 0 and 1000.
```

For a trajectory or multiple points, append exactly:

```text
Return only a JSON list like [{"point_2d": [x1, y1], "label": "point_1"}, {"point_2d": [x2, y2], "label": "point_2"}, ...]. Use integer coordinates between 0 and 1000.
```

## Training

The vision encoder is frozen and the vision projector is trainable. The model was fully fine-tuned for one epoch with a learning rate of 2e-5 and a maximum sequence length of 5600.

## Evaluation

This is the paper's primary 4B model. The paper reports a 62.7 pointing/VQA average for this model. The local full benchmark evaluation associates the released weights with aggregate score 0.698.

| Model | Aggregate | Where2Place | RefSpatial location | IA-Bench | RoboRefIt testA | VA Bench-P |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| Qwen3.5-4B | 0.698 | 72.0 | 59.0 | 79.0 | 85.7 | 65.7 |
| Qwen3.5-0.8B-VTFT | 0.605 | 58.0 | 47.0 | 76.7 | 80.9 | 48.3 |
| Qwen3.5-9B-EO | 0.719 | 76.0 | 68.0 | 78.5 | 85.2 | 68.7 |

## 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}
}
```