|
--- |
|
base_model: |
|
- ToastyPigeon/ms3-roselily-instruct |
|
library_name: transformers |
|
tags: |
|
- mergekit |
|
- merge |
|
|
|
--- |
|
# todo |
|
|
|
make a model card and put a cute girl on it |
|
|
|
# some info |
|
|
|
Making this public so it can be tried and possibly merged if desired while I work on getting the energy to write a proper card. |
|
|
|
Short list of things to know: |
|
- This is a bunch of RP, story writing, etc. creative data applied to [ToastyPigeon/ms3-roselily-instruct](https://huggingface.co/ToastyPigeon/ms3-roselily-instruct). |
|
- Instruct format: ChatML or Alpaca preferred, Tekken v7 possible |
|
- ChatML tokens were assigned to unused tokens 20 and 21, this leaves all the tekken tokens intact so merges w/ tekken models are feasible |
|
- Instruct-tuning phase did include Tekken v7 so the tokens are initialized and recognized, but I did not continue with it on the creative step because I do not like it for creative stuff (too restrictive with turn order) |
|
- Feels a little less sensitive to samplers than Instruct-based MS3 models, but should probably still be used with conservative samplers |
|
|
|
# chat templates |
|
|
|
You may need to set `<|im_end|>` and/or `</s>` as stopping strings depending on which format you're using, the model generates both properly but tokenizers can be finicky about what they stop on by default |
|
|
|
Alpaca w/ System |
|
``` |
|
### System: |
|
{system prompt} |
|
|
|
### Instruction: |
|
{user message} |
|
|
|
### Response: |
|
{model answer}</s> |
|
``` |
|
ChatML |
|
``` |
|
<|im_start|>system |
|
{system prompt}<|im_end|> |
|
<|im_start|>user |
|
{user message}<|im_end|> |
|
<|im_start|>assistant |
|
{model answer}<|im_end|> |
|
``` |
|
Also saw some completion training in chat mode and adventure mode. |
