---
pretty_name: ContrastGround
configs:
- config_name: corpus
  data_files:
  - split: train
    path: data/corpus/train.jsonl
- config_name: selectground-8b
  data_files:
  - split: train_pairs
    path: data/selectground-8b/train_pairs.jsonl
  - split: train_replay
    path: data/selectground-8b/train_replays.jsonl
- config_name: selectground-30b-a3b
  data_files:
  - split: train_pairs
    path: data/selectground-30b-a3b/train_pairs.jsonl
  - split: train_replay
    path: data/selectground-30b-a3b/train_replays.jsonl
task_categories:
- image-to-text
tags:
- gui-grounding
- computer-use
- contrastive-learning
- sft
license: other
---

# ContrastGround

ContrastGround is the training corpus for the updated SelectGround-8B and
SelectGround-30B-A3B checkpoints. It combines UI-Vision-cleaned ClickContrast
examples with teacher-verified hard negatives mined from GroundCUA. It does not
contain UI-Vision examples.

## Which configuration should I use?

| Configuration | Pair rows | Replay rows | Released model |
|---|---:|---:|---|
| `selectground-8b` | 4,942 | 4,103 | `ruotian/SelectGround-8B` |
| `selectground-30b-a3b` | 4,544 | 3,824 | `ruotian/SelectGround-30B-A3B` |
| `corpus` | 4,944 | 4,103 | Auditable union of both recipes |

The two model configurations are intentionally different. Use the exact
configuration named by the model card to reproduce that checkpoint. The
`corpus` rows include `training_configs`, which records membership in one or
both released recipes.

```python
from datasets import load_dataset

dataset = load_dataset("ruotian/ContrastGround", "selectground-8b")
pairs = dataset["train_pairs"]
replay = dataset["train_replay"]
```

The `image` field is a repository-relative content-addressed path. When using a
snapshot checkout, resolve it relative to the snapshot root.

## Schema

Every row contains `image`, `image_width`, `image_height`, `instruction`,
`response`, `target_bbox`, `source_ref`, `role`, and `source_license`. Pair rows
also contain `distractor_bbox` and `candidate_bboxes`, which supervise the
SelectGround auxiliary selection loss. Coordinates in bounding boxes are
absolute source-image pixels; `response` is the target center in normalized
0–1000 coordinates.

## Construction

The ClickContrast component starts from Click-100k and removes all examples
whose canonical source is UI-Vision (`element_grounding` or
`layout_grounding`). Its replay remains stratified across the retained
Click-100k sources.

The GroundCUA component uses two natural-language views per target. A target is
eligible only when MAI-UI-8B and UI-Venus-1.5-8B both land inside the original
pixel-space ground-truth box, while Qwen3-VL-8B misses the expanded target
neighborhood and lands on a different native UI element. Spatial prompts also
require a unique text anchor and an unambiguous nearest target. The released
views differ only in how conservatively those GroundCUA pairs and replay rows
are sampled.

Pinned upstream revisions and SHA-256 checksums for every JSONL are recorded in
`manifest.json`. Images are named by their byte-level SHA-256 digest; 8,423
source paths collapse to 6,834 distinct image files.

## Licensing and provenance

GroundCUA is distributed under MIT by ServiceNow. At the pinned revision used
here, the Click-100k dataset card does not declare a dataset-level license.
Accordingly this mixed-source dataset uses `license: other`, and every row
records its source and source-license status. Users are responsible for
checking the upstream terms for their intended use. This card does not
relicense upstream images or annotations.

The optional teacher-point provenance fields inside `source_ref` are
JSON-encoded coordinate strings so that Click-100k and GroundCUA rows share a
stable Arrow schema; they are not training targets.

## Evaluation leakage

The ClickContrast component was rebuilt to exclude UI-Vision provenance before
training. ScreenSpot-Pro, UI-Vision, and OSWorld-G are not validation splits of
this dataset. Reported released checkpoints were selected using those public
benchmarks, so their scores should be treated as test-tuned rather than as
held-out model-selection estimates.