Crisis and resilience · Shipped
crisis-agent, built to work with the network down
A fine-tuned 7B assistant for crisis guidance, published in full and in GGUF form so it runs on a laptop with no connection.
HuggingFace, GGUF, Python ·
What it found
The quantised offline build mattered more than the accuracy gain. In an emergency the network goes first, so a slightly worse model that runs locally beats a better one that cannot be reached.
- 5models published
- 7Bparameters
The work
What was actually built
A crisis assistant that needs an internet connection has a problem, because the connection is frequently the first thing a crisis takes away.
Two generations exist. crisis-agent-v1 established that a 7B model fine-tuned on the dataset above could hold a useful conversation about emergency preparation. crisis-agent-v2 improved it. Both are published twice: once as weights for anybody who wants to continue the work, and once as GGUF, which is the form that runs on consumer hardware without a server.
The GGUF build was initially an afterthought, a convenience for local testing. It turned out to be the point.
Still open
What this did not settle.
How much does quantisation cost on advice that matters?
The GGUF builds were checked for coherence rather than measured against the full-precision weights on crisis-specific questions. That comparison is the obvious next piece of work and it has not been done.
See it
Open it, or read the code.
huggingface.co. Opens in a new tab on an external site.