kilterbraingen

Generating kilter board problems with transformers

Usage

Clone the repo

git clone https://github.com/rparrett/kilter_brain_gen
cd kilter_brain_gen

Install dependencies

We're using uv for package and project management.

• Run uv sync
• (Windows / CUDA) (TODO, this might be configured correctly in pyproject.toml)

Get climb data

• Install sqlite
• Use boardlib, or extract the database from a kilter apk file, or ask someone who has it.
• Move that database to data/climbs.sqlite3
• Run uv run src/data/get_csv.py

Train the climb_gpt model

uv run src/climb_gpt/train.py

Generate some climbs

uv run src/climb_gpt/generate.py

Run the API server

• uv run flask -A src/api/api.py run
• Test with curl -X POST -H "Content-Type: application/json" -d '{"prompt":""}' 'http://127.0.0.1:5000/generate'
• Debug builds of kilter_brain will connect to the local server.

Development

Linting

uv run ruff check

Formatting

uv run ruff format

Monitoring Experiments

uv run tensorboard --logdir models

TODO

• Train a model for route names that actually works
• Tidy everything up with a nice CLI framework
• Add similarity search
• Somehow specify windows-specific dependencies that grab torch with CUDA support

Experiments

• Try adding duplicate climbs with randomized frame data (shuffle-frames)
• Try different sampling strategies for climb data