A GitHub repository lets users train and chat with a frozen language model. The model is coupled to the full MaleCNS v1.0 fly connectome. It uses a 278,528-parameter adapter to read the graph’s state. The graph has 166,700 nodes and 25,582,938 directed connections. Only the adapter is trained. The language ability comes from a pretrained Liquid AI model. The repository supports local training and inference without API keys. It requires Python 3.12 on macOS or Linux. Users need at least 10 GB of free disk space. 16 GB or more RAM is recommended. CPU training is supported but slower. The code is open-source under MIT license. The paper linked does not claim fly anatomy improves performance.
What it unlocks
Training and chatting with a language model coupled to a full fly brain connectome.
- 166,700 retained nodes
- 25,582,938 directed connections
- 278,528-parameter adapter
What you need to act on it
- Python 3.12
- macOS or Linux
- 10 GB+ free disk space
- 16 GB+ RAM preferred
Sources