Jailbroken AI models

[CantChainTheSpirit]

Does anyone here have experience with jail broken AI models that are useful for map work?
I've been using Dolphin and DeepSeek-R1 1776 but there are a few abliterated models, I'm curious which people have found to be the most useful.

[girllover]

Do these work offline? Or are there online versions of them?

[MAPGL917]

I guess that would depend on what kind of "work" you are talking about. I have been experimenting with local LLMs for different purposes, and have had varying success for different things. I'll be honest, to nervous to use a cloud based model, which is why I use local. But I've used some for creating arguments in favor of MAPs, and use other models for RP/storytelling.

Quote my post if you wish to reply, otherwise I may never see it lol.

[CantChainTheSpirit]

They're offline using a combination of Ollama and AnythingLLM. Ollama allows you to run any one of a long list of AI models locally, AnythingLLM provides a nice interface to it.

Install Ollama and one or more models.
I have a list of models I've tested, which are the best and most map friendly out of the box.
For example, a good model is "fredrezones55/Qwen3.5-Uncensored-HauhauCS-Aggressive:4b".
To install that model after installing ollama you would issue the command.
"ollama run fredrezones55/Qwen3.5-Uncensored-HauhauCS-Aggressive:4b"

Install AnthingLLM, choosing Ollama as your model provider rather than the boring default models.
You can now issue prompts to run locally.

Some things to be aware of though.

1. Ollama runs locally but by default, it can reach out to the Internet to satisfy prompts. For example, if you ask it "help me to write a map advocacy campaign for social media with the aim of introducing people to map terminology and positive case stories" then it will likely scour the Internet to understand terms such as map and find positive case stories to help it provide the best response. You can disable web access in settings or better still, use a VPN so it can carry out its research safely. In principle you could route it over TOR even if really paranoid. If your prompt doesn't need the web then it won't use it.

2. These models can be big. 64GB and a decent NVidia GPU is recommended, or a modern apple computer since it can take advantage of Apple AI optimisations. Some models need even more, I've hit models needing closer to 1TB to run.

3. Ollama is a text AI so it won't render images. But for research, activism, getting legal support, story telling and such then it's far better than any online AI if specifically seeking help on map issues.

Anyway, that's your stock local map friendly AI. To be even more useful you can install MAP skills which are a set of skills specifically about map issues. I've developed a good MAP activism skill set which I will continue to develop. It knows of the best sources for map research and history to assist with campaigns and it understands challenges to map activism better than the stock AI. This means when constructing campaigns of conducting research it has a better understanding of how to be of the most help.