FLM code repository (github.com/nftechie/flm)

The FLM repository on GitHub is a local training and chat setup built around an odd pair: a frozen language model and a fly connectome. The graph comes from MaleCNS v1.0, with 166,700 retained nodes and 25,582,938 directed connections. Token embeddings feed the whole fixed graph, and a small 278,528-parameter adapter reads that state to nudge the next-token scores of Liquid AI’s LFM2.5-1.2B-Instruct.

Only the adapter learns. The language ability comes from the pretrained backbone, not from any claim that a fly understands English, which is the kind of sentence the project has to say out loud because people will otherwise try to turn anatomy into mysticism. The repository stays entirely local: no API key, no account, no hosted inference service, no web server.

Setup is straightforward, if a little hardware-dependent. It targets Python 3.12 on macOS or Linux, works on Apple Silicon through MPS, on NVIDIA GPUs when the installed PyTorch build supports CUDA, and on CPU as a fallback. The project asks for several gigabytes of downloads, at least 10 GB of free disk space, and preferably 16 GB or more of RAM. CPU training works, just more slowly.

The training scripts are laid out step by step, from cloning the repo to building the graph, optionally compiling a C kernel, and running conversation training. Downloads are pinned to upstream revisions and the original connectome files are checked against SHA-256 hashes. If there’s no compiler, the SciPy path still does the same recurrence; the C kernel is only an accelerator for feature extraction, not a different model.

There’s a real effort to make the results inspectable rather than theatrical. Training uses 64 corpus conversations and 32 synthetic style examples, with validation choosing the checkpoint and 24 separate test conversations used afterward. A direct-input adapter with the same parameter count is trained as a control, and the repo is careful to say the matched control did slightly better in the separate frozen study described in the paper. That is a tidy little rebuke to anyone hoping the fly anatomy secretly won the day.