Get started

From install to your first names

You need Linux on x86_64 and an NVIDIA GPU with its driver. The installer puts CausalBridge in its own environment and does not touch your Python.

01/ 07

Install

Downloads CausalBridge and installs the causalbridge command in its own environment.

curl -fsSL https://raw.githubusercontent.com/Sullivan07043/CausalBridge/main/install.sh | sh
Terminal
$ curl -fsSL https://raw.githubusercontent.com/Sullivan07043/CausalBridge/main/install.sh | sh
Getting causalbridge-0.2.2-cp312-cp312-linux_x86_64.whl
Checking the wheel against the published sha256
Creating the environment (Python 3.12)
Installing CausalBridge 0.2.2 and its dependencies (PyTorch is large)

Installed CausalBridge 0.2.2: /home/you/.local/bin/causalbridge

Next steps:
  1. Download a language model, for example Qwen/Qwen3-4B-Instruct-2507:
       /home/you/.local/lib/causalbridge/current/bin/hf download Qwen/Qwen3-4B-Instruct-2507
     A model on your GPU supports every feature.
     A hosted API supports text mode only, without the causal prefix.
  2. Start CausalBridge and open it in your browser:
       causalbridge activate
     Open it again: causalbridge gui    Stop it: causalbridge exit
  Command line: causalbridge --help
02/ 07

Get a language model

Download the model you want. CausalBridge never downloads one for you. A model on your GPU supports every feature. A hosted API supports text mode only, without the causal prefix.

~/.local/lib/causalbridge/current/bin/hf download Qwen/Qwen3-4B-Instruct-2507
Terminal
$ ~/.local/lib/causalbridge/current/bin/hf download Qwen/Qwen3-4B-Instruct-2507
Fetching 13 files: 100%|██████████| 13/13
/home/you/.cache/huggingface/hub/models--Qwen--Qwen3-4B-Instruct-2507/snapshots/cdbee75f17c01a7cc42f958dc650907174af0554
03/ 07

Activate

Starts CausalBridge and opens it in your browser. If you close the browser, causalbridge gui opens it again. causalbridge exit stops CausalBridge.

causalbridge activate

Over ssh, forward the port from your computer first:

ssh -L 8765:127.0.0.1:8765 <user>@<gpu machine>
Terminal
$ causalbridge activate
CausalBridge 0.2.2 is running.
Address: http://127.0.0.1:8765/?token=...
Opened it in your browser.
Open it again: causalbridge gui    Stop it: causalbridge exit

# the browser was closed
$ causalbridge gui
Address: http://127.0.0.1:8765/?token=...
Opened it in your browser.
Open it again: causalbridge gui    Stop it: causalbridge exit

$ causalbridge exit
CausalBridge stopped.
04/ 07

Choose the data

Choose the data file and say whether each row is an independent sample or a time step. The page shows the first rows.

127.0.0.1:8765
The Data card: the data file, the kind of rows and a preview of the first rows
05/ 07

Give the names you know

Type the names you know, or load a names file. CausalBridge names every column that you leave empty.

127.0.0.1:8765
The Names card: a table of columns, ten with names and two empty
06/ 07

Choose the graph and the model

Discover the graph from the data, or give your own graph file. Choose a model on your GPU or a hosted API, then run.

127.0.0.1:8765
The Graph and Language model cards
07/ 07

Train for your domain

Choose a folder of fully named datasets. CausalBridge checks the folder and trains a profile on your GPU. Training continues when you close the browser.

127.0.0.1:8765
The Train page with a checked folder of four datasets

Every page of the interface shows the equivalent command. The manual describes all commands for use without the interface.

License

CausalBridge is proprietary software. It is free for research and evaluation. Commercial use needs a separate written license from the CausalBridge Team.

Read the license · Ask about a commercial license · Third-party notices

Citation

If you publish results obtained with CausalBridge, cite the CausalBridge paper, How Causality Bridges the Semantic Gap.

@misc{zhang2026causalitybridgessemanticgap,
  title={How Causality Bridges the Semantic Gap},
  author={Shuhao Zhang and Xuran Zhou and Han Guo and Pengtao Xie and Yujia Zheng},
  year={2026},
  eprint={2610.02594},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2610.02594},
}