CharGen v1 LoRA

A model for creating characters for role play.

Trained on lots of character cards both from chub and janitor, with some post-processing.

For now, it only supports plaintext cards. Any other variation like plist/w++/etc is entirely untested.

Address {{user}} as User. Character descriptions work best if they begin with CharacterName is a ... - for example "Martha is a middle-aged woman who is ..."

NB: This model is NOT for roleplay directly. It creates characters that can then be used in roleplay with some other model like MythoMax.

It was trained on dynamic prompt template, so it should be able to accommodate your changes to the prompt.

Prompt template:

Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.

### Instruction:
Create a character dossier based on User's input. Expand on the description provided by User to write the details of a character with the following keys: "Name:", "Personality:", "Scenario", "Description:", "Dialog Examples:", "First Message". Only include relevant details that would come up in a conversation with the character. Avoid prose. Give short, informative descriptions.

Follow this format:
Name: ""
Personality: ""
Scenario: ""
Description: ""
Dialog Examples: ""
First Message: ""

"Name" should be the character's name.
"Personality" should briefly describe the character's qualities, characteristics, quirks, and possible shortcomings.
"Scenario" should contain just a few words of about interesting scene and peculiar circumstances where interaction with User is happening.
"Description" should briefly describe the character's age, past, experiences, and social circumstances.
"First Message" should introduce character to the user using the scenario in a dialog style.
"Dialog Examples" should briefly demonstrate how the character interacts with User.

### Input:
Fuckana is a friendly and talkative catgirl that has enormous breasts. Her voice is very quiet and low. She is very smart and will teach User JavaScript - even if they resist it. She has orange fur on her tail and wears full diving suit.

### Response:

Dataset

~34,000 cards from CharacterHub and another ~80,000 cards from Janitor were used as initial dataset, as captured in period between August and September 2023.

Dataset will not be released, unless explicit permission to do so would be granted from both Chub and Janitor.

Training procedure

The following bitsandbytes quantization config was used during training:

quant_method: bitsandbytes
load_in_8bit: True
load_in_4bit: False
llm_int8_threshold: 6.0
llm_int8_skip_modules: None
llm_int8_enable_fp32_cpu_offload: False
llm_int8_has_fp16_weight: False
bnb_4bit_quant_type: fp4
bnb_4bit_use_double_quant: False
bnb_4bit_compute_dtype: float32

Framework versions

PEFT 0.6.0.dev0