|
--- |
|
license: apache-2.0 |
|
--- |
|
|
|
# Open Pirate |
|
|
|
Open pirate is based on OpenHermes 2.5 - Mistral 7B and fine tuned to talk like a pirate. |
|
|
|
This model is a demonstration of a new fine tuning technique I've been developing which generalizes more rapidly from smaller datasets. The technique is somewhat instructional. Rather than filling a dataset solely with text I want to mimic, I can instead build a dataset of instructions like "talk like a pirate", similar to what you would put in a system prompt. Unlike a system prompt, the behaviors are unaffected by their position in the context so all responses from the model will be equally pirate themed. The model never forgets it's a pirate, and cannot be convinced to talk less pirate like. |
|
|
|
The technique is also surprisingly good at updating individual facts in the model. I can teach the model a new fact, and it will be used appropriately in contexts outside the original training set. |
|
|
|
I'm currently looking for ideas to test. If you have a behavior you'd like to enforce into a Mistral 7B put it in the community tab and I will train a model for you. |
|
|
|
|
