---
pretty_name: ConClick-Data
license: other
license_name: upstream-source-terms
license_link: https://huggingface.co/datasets/mlfoundations/Click-100k
task_categories:
- image-to-text
tags:
- gui-grounding
- computer-use
- contrastive-learning
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: all_pairs
    path: data/all_pairs.jsonl
  - split: all_replay
    path: data/all_replay.jsonl
---

# ConClick-Data

Training data for [ConClick](https://github.com/zhangruotian/ConClick): **4,292 target–distractor pairs + 4,292 ordinary grounding examples** from [Click-100k](https://huggingface.co/datasets/mlfoundations/Click-100k), covering 6,690 screenshots. 

| File | Examples | Supervision |
|---|---:|---|
| `data/all_pairs.jsonl` | 4,292 | Target coordinate, target/distractor boxes, candidate boxes |
| `data/all_replay.jsonl` | 4,292 | Target coordinate |

Download with screenshots, preserving relative paths:

~~~bash
hf download ruotian/ConClick-Data --repo-type dataset --revision conclick --local-dir data/conclick-data
~~~

Rows include `id`, `image`, `instruction`, `response`, image dimensions/checksum, and source provenance. Responses use 0–1000 coordinates; boxes use `[left, top, right, bottom]` in original-image pixels. Training instructions and [ConClick-8B](https://huggingface.co/ruotian/ConClick-8B) are linked from the code repository.

Screenshots retain the terms of Click-100k and its constituent sources; the code's Apache-2.0 license does not relicense the data.