Document release-ready filtered SPARC VQA split

README.md

@@ -0,0 +1,84 @@
+---
+pretty_name: SPARC VQA
+tags:
+  - robotics
+  - vision-language
+  - spatial-reasoning
+  - sparc
+---
+
+# SPARC VQA
+
+SPARC VQA is the generated spatial VQA training dataset used in the SPARC Qwen3.5 model releases. Each example embeds its image bytes and includes a question, answer, task type, target type, source dataset identifier, split, and JSON metadata.
+
+## Ready-to-train split
+
+Use `train_filtered_t097_mpo700.parquet` for SPARC-only training. It contains the exact 284,909 examples retained by the release configuration, with no further SPARC filtering required:
+
+- `selected_start_score >= 0.97`
+- sorted by `selected_start_score` in descending order before diversity selection
+- a maximum of 700 `(source, trajectory_name, subtask_index)` groups per normalized object name
+- vacant-goal and vacant-start examples mentioning `gripper` excluded
+
+The unfiltered 838,211-example source corpus is available at [irl-kit/SPARC-VQA-Raw](https://huggingface.co/datasets/irl-kit/SPARC-VQA-Raw). It includes `export_sparc_training_subset.py` for reproducing this split.
+
+## 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 |
+
+## Mixtures
+
+| Release | Data mixture |
+| --- | --- |
+| Qwen3.5-4B | SPARC VQA + FSD + RoboPoint + LLaVA-OneVision2 |
+| Qwen3.5-0.8B-VTFT | SPARC VQA + FSD + RoboPoint + LLaVA-OneVision2 |
+| Qwen3.5-9B-EO | SPARC VQA + FSD + RoboPoint + LLaVA-OneVision2 + EO-1.5M |
+
+FSD, RoboPoint, LLaVA-OneVision2, and EO-1.5M are external datasets and should be retrieved from their upstream releases under their respective terms. The released models are [Qwen3.5-4B](https://huggingface.co/irl-kit/SPARC-Qwen3.5-4B), [Qwen3.5-0.8B-VTFT](https://huggingface.co/irl-kit/SPARC-Qwen3.5-0.8B-VTFT), and [Qwen3.5-9B-EO](https://huggingface.co/irl-kit/SPARC-Qwen3.5-9B-EO).
+
+## Reproducing the SPARC filter
+
+The included exporter applies the exact release filtering logic without decoding images during its selection pass:
+
+```bash
+python export_sparc_training_subset.py \
+  /path/to/train.parquet \
+  train_filtered_t097_mpo700.parquet \
+  --quality-threshold 0.97 \
+  --max-per-object 700 \
+  --blocked-vacant-location gripper
+```
+
+The raw dataset repository includes the accompanying `ours_adaptive_det_soft_snr_sp8.yaml` generation configuration used by the training runs.
+
+## Prompting compatibility
+
+The released models are sensitive to output formatting. Use each model's bundled chat template with a single user turn containing the image followed by the question. For point prediction, use this suffix:
+
+```text
+Output the point coordinates in JSON format like [{"point_2d": [x, y], "label": "target"}]. Use integer coordinates between 0 and 1000.
+```
+
+For trajectories or multiple points, use this suffix:
+
+```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.
+```
+
+## 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}
+}
+```