|
--- |
|
license: apache-2.0 |
|
base_model: |
|
- mistralai/Mistral-Nemo-Instruct-2407 |
|
--- |
|
|
|
# Sto'Vo'Kor 12B |
|
|
|
[mistralai/Mistral-Nemo-Instruct-2407](https://huggingface.co/mistralai/Mistral-Nemo-Instruct-2407) fine-tuned on a private collection of ~30M tokens worth of instruct and multi-turn RP. |
|
|
|
*Note: this model is an experiment separate from the Unslop Initiative. You will probably get whispers and spine shivers with this model. I think it's pretty good anyway.* |
|
|
|
## Instruct Format |
|
|
|
Instruct format is V3-Tekken, the same as Mistral Nemo Instruct (except the chat template used won't freak out if your turns get mixed up, like tends to happen in ST. Thanks, fizz!) |
|
|
|
``` |
|
<s>[INST]{System or user instructions}[/INST]{AI Response}</s> |
|
``` |
|
During training, system turns were given as the first user turn in the conversation, separate from the user character's first turn. i.e., System as user -> AI turn (filler or first turn) -> User first turn |
|
|
|
## Recommended Samplers |
|
|
|
Whatever you're used to for Nemo should work. For me this is stable with: |
|
- temp 0.7 |
|
- min-p 0.03 |
|
- DRY 0.5/1.75/5/1024 |
