ruotian/ClickContrast

ClickContrast

ClickContrast contains 4,292 target–distractor pairs and 4,292 coverage examples from Click-100k, with 6,690 distinct screenshots. It is the training dataset for SelectGround-8B.

Training uses all_pairs and all_replay with a 1:1 pair-to-coverage sampling cycle.

Pairs retain teacher-supported targets together with a model prediction on another parsed element outside a centered 2x target neighborhood. A focused subset covers teacher-checked small controls and icons. Coverage examples span 86 Click100k source categories.

Format

Download the entire repository to retain the relative screenshot paths:

~~~bash hf download ruotian/ClickContrast --repo-type dataset --revision paper --local-dir data/clickcontrast ~~~

Each row contains:

• —id: stable example identifier.
• —image: screenshot path relative to this repository.
• —image_width, image_height, image_sha256: original dimensions and file checksum.
• —instruction: original grounding instruction.
• —response: target center as [x,y], with integer coordinates normalized to 0–1000.
• —source: upstream dataset, split, and row index.

Paired rows additionally contain target_bbox, distractor_bbox, and candidate_bboxes, all in original-image pixels with [left,top,right,bottom] order. The JSONL order is part of the frozen training recipe. The manifest records file hashes and counts.

Training

Use the SFT-plus-selection-loss recipe in the SelectGround code repository. All examples contribute coordinate loss. Only paired examples contribute selection loss. No external parser or teacher is needed to train from this release.

Source terms

Screenshots are a subset of Click-100k and retain their upstream provenance. They are not relicensed as project code. Use is subject to the applicable terms of Click-100k and its constituent sources. The repository's code license does not replace those data terms.