WebUniverse-NSL/Qwen3.5-WebUniverse
Image-Text-to-Text • 35B • Updated • 5
Datasets:
Access WebUniverse-Mix
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
By requesting access, you agree to the WebUniverse-Mix License and the applicable upstream terms for the portions you use. The release is provided as is, without warranty.
Log in or Sign Up to review the conditions and access this dataset content.
WebUniverse-Mix
15,000 verified development trajectories for training web-development agents from written requirements and visual references. This is the main training set for Qwen3.5-WebUniverse, with a 70:30 text-grounded to vision-grounded mixture.
- 🌐 Project Website — Coming soon
- 📄 Paper — Coming soon
- 💻 GitHub Repo — Coming soon
- 🤗 HF Datasets:
- WebUniverse-Mix — 15,000 training trajectories. [You are here!]
- WebUniverse-Env — 15,000 task environments.
- 🤖 HF Models — Main model and the StepFun-data comparison model.
Composition
| Task variant | Trajectories |
|---|---|
| Text-grounded / standalone | 6,355 |
| Text-grounded / iterative | 4,145 |
| Vision-grounded / single-page | 1,283 |
| Vision-grounded / multi-page | 1,069 |
| Vision-grounded / multi-viewport | 2,148 |
| Total | 15,000 |
The exact 5K and 10K ablation subsets are provided as ordered ID lists. See subset loading.
Usage
During release preparation, downloads require manual approval. After approval, run hf auth login.
from datasets import load_dataset
ds = load_dataset("WebUniverse-NSL/WebUniverse-Mix", split="train", streaming=True)
row = next(iter(ds))
Images are embedded in each row as {bytes, path}. Detailed usage covers loading and metadata. Use export_ms_swift.py to export JSONL and images for ms-swift training.
Format
Each Parquet row contains a trajectory, its images, and metadata.
| Field | Contents |
|---|---|
messages |
Ordered messages with role, content, and loss |
images |
Image bytes and paths, in trajectory order |
system |
System prompt |
tools |
JSON-serialized tool definitions |
id |
Stable trajectory ID |
task_family, variant |
Task category |
collection_mode, refinement_type, trajectory_model |
Collection metadata |
line, n_tokens, n_tool_calls, n_images, observed_framework |
Row index and trajectory statistics |
The release preserves message text and image bytes; it is not tokenized or packed for a specific model. Message roles are user, assistant, tool_call, and tool_response. loss: false masks a message from supervision; loss: null uses the training framework's default. Preserve these masks when exporting data.
Environments
WebUniverse-Env contains the task environments for these trajectories.
Source seeds
- Text seeds — Text-task seeds and source links.
- Visual task sources — Source websites for each visual task.
License
Released under the WebUniverse-Mix License.
- Downloads last month
- 115
Models trained or fine-tuned on WebUniverse-NSL/WebUniverse-Mix