Getting started

This project separates model-free repository checks from optional model-backed runs. Start with the first path; it is fast and does not download weights.

1. Install the local tools

Install uv and Node.js 20 or newer if you want to build the docs site locally. Python commands in this repository should run through uv.

git clone https://github.com/Sunwood-ai-labs/jev-colab-lab.git
Set-Location jev-colab-lab
uv --version

2. Run model-free checks

uv run --no-project --with pytest pytest experiments/laya/tests experiments/openjev-nli/tests experiments/semif/tests -q
uv run --project experiments/jevlike --extra dev pytest experiments/jevlike/tests -q
uv run --no-project python -m py_compile experiments/kev/t4_inference.py

These tests validate helper logic, public fixtures, notebook structure, and Python syntax. They do not prove that a GPU is available or that a model download will succeed.

3. Choose an experiment

Open the corresponding experiment README and check its source manifest before installing anything. The experiment README is the authority for its own dependency constraints and result file.

Target	Experiments	Intended use
T4	Laya, Kev-0.5B, Jevlike	lower-cost decision-model and tiny scorer paths
L4	SemIf 4B, OpenJev NLI 4B	4B direct-logit and NLI paths

4. Run on Colab from Windows

The official Google Colab CLI currently supports Linux and macOS, not Windows. Use WSL Ubuntu on a Windows host:

uv tool install google-colab-cli
colab --help

Each run should have:

• a unique session name such as jev-laya or jev-semif
• a task-specific session state file outside the repository
• the requested GPU recorded separately from the actual GPU reported in the result
• an explicit download of sanitized results before stopping a manually created session

Authentication, quota, and entitlement failures are valid outcomes to record. Do not commit OAuth links, ADC files, session logs, or weights.

5. Read the evidence

Open the linked result JSON after a run. Treat the GPU as successful only when the result has a successful status and the actual GPU is recorded. See Reproducibility for the measurement vocabulary.