Datasets:
File size: 2,847 Bytes
01ba272 2651c95 01ba272 2651c95 01ba272 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 | ---
pretty_name: Jev GUI Training Preview
license: other
language:
- en
task_categories:
- image-text-to-text
tags:
- gui-agent
- multimodal
- decision-model
- jev
- qwen
size_categories:
- 100K<n<1M
configs:
- config_name: preview
data_files:
- split: train
path: preview/train.parquet
- config_name: decisions
data_files:
- split: train
path: decisions/train-*.parquet
- config_name: images
data_files:
- split: train
path: images/train-*.parquet
---
# Jev GUI Training Preview
This public release contains the GUI-only weighted training mixture prepared
for Jev-Vision:
- **636,800 decision references**
- **132,647 unique images**, stored once
- Choice, yes/no (`noul`) and ordered-score decisions
- Single-step, short-horizon, teacher-generated and UI-element tasks
The weighted sampler intentionally repeats some records. `sample_index` is
unique, while `id` identifies the underlying decision and may repeat.
## Configurations
### `preview`
A small image gallery for the Hub dataset viewer. The complete normalized
tables are in `decisions` and `images`.
### `decisions`
One row per sampled training reference. `image_ids` contains ordered content
hashes referring to rows in the `images` configuration. `state_json` and
`meta_json` preserve structured fields as JSON strings for a stable Parquet
schema.
### `images`
One row per unique image. The `id` column is the content hash used by
`decisions.image_ids`; `image` is a Hugging Face Image feature.
```python
from datasets import load_dataset
decisions = load_dataset(
"IntelligenceLab/Jev-GUI-Training-Preview",
"decisions",
split="train",
)
images = load_dataset(
"IntelligenceLab/Jev-GUI-Training-Preview",
"images",
split="train",
)
```
## Sample composition
| Source | Sampled rows |
|---|---:|
| `aguvis` | 180,087 |
| `aguvis_short_horizon` | 213,646 |
| `tasksource_vision` | 4,967 |
| `teacher_gui` | 126,723 |
| `teacher_stageb` | 51,390 |
| `waveui` | 59,987 |
The exact weighted mixture is in `training_mix_gui_v1.json`; build metadata is
in `manifest_meta.json` and `release_stats.json`.
## Important rights and privacy notice
This release has mixed and unresolved source-image rights and may include
personal-looking text present in public GUI traces. Read `LICENSES.md` before
use. `license: other` does not grant rights beyond each upstream source.
## Provenance
- [AGUVIS Stage 2](https://huggingface.co/datasets/xlangai/aguvis-stage2)
- [Wave-UI 25K](https://huggingface.co/datasets/agentsea/wave-ui-25k)
- [Tasksource Jev Typed Decisions](https://huggingface.co/datasets/tasksource/tasksource-jev-typed-decisions)
- Teacher-generated GUI decisions from Qwen3.5-122B-A10B
- Deterministic seed: `20261027`
- Mix configuration SHA-256: `632255ca91150dd16e46f0717bb2433866d38bc5d9a48656ba1aa3d07c6b483a`
|