Decensored using a custom training script guided by activations, similar to ablation/"abliteration" scripts but not exactly the same approach.
The training script is released under the MIT license:
https://github.com/nkpz/DeLMATThe base model says it's for finance, but in my experiences, it's a very good all-purpose reasoning model. Shockingly easy to tune - this was my first try, and there's nothing I would change about it. It does exactly what it's told with a detailed and often entertaining thought process.
