spow12/ChatWaifu_v2.0_22B

Model Card for Model ID

Merged model using mergekit

This model aimed to act like visual novel character.

Merge Format

models:
  - model: mistralai/Mistral-Small-Instruct-2409_sft_kto
    layer_range: [0, 56]
  - model: mistralai/Mistral-Small-Instruct-2409
    layer_range: [0, 56]
merge_method: slerp
base_model: mistralai/Mistral-Small-Instruct-2409_sft_kto
parameters:
  t:
    - filter: self_attn
      value: [0, 0.5, 0.3, 0.7, 1]
    - filter: mlp
      value: [1, 0.5, 0.7, 0.3, 0]
    - value: 0.5 # fallback for rest of tensors
dtype: bfloat16

WaifuModel Collections

• TTS
• Chat
• ASR

Unified demo

WaifuAssistant

Update

• 2024.10.11 Update 12B and 22B Ver 2.0
• 2024.09.23 Update 22B, Ver 2.0_preview

Model Details

Model Description

• Developed by: spow12(yw_nam)
• Shared by : spow12(yw_nam)
• Model type: CausalLM
• Language(s) (NLP): japanese, english
• Finetuned from model : mistralai/Mistral-Small-Instruct-2409

Currently, chatbot has below personality.

character	visual_novel
ムラサメ	Senren＊Banka
茉子	Senren＊Banka
芳乃	Senren＊Banka
レナ	Senren＊Banka
千咲	Senren＊Banka
芦花	Senren＊Banka
愛衣	Café Stella and the Reaper's Butterflies
栞那	Café Stella and the Reaper's Butterflies
ナツメ	Café Stella and the Reaper's Butterflies
希	Café Stella and the Reaper's Butterflies
涼音	Café Stella and the Reaper's Butterflies
あやせ	Riddle Joker
七海	Riddle Joker
羽月	Riddle Joker
茉優	Riddle Joker
小春	Riddle Joker

Chat Format

<s>This is another system prompt.
[INST]
Your instructions placed here.[/INST]
[INST]
The model's response will be here.[/INST]

Usage

You can use above chara like this

from huggingface_hub import hf_hub_download
hf_hub_download(repo_id="spow12/ChatWaifu_v1.2", filename="system_dict.json", local_dir='./')

with open('./system_dict.json', 'r') as f:
    chara_background_dict = json.load(f)

chara = '七海'
background = chara_background_dict[chara]
guideline = """
Guidelines for Response:
Diverse Expression: Avoid repeating the same phrases or reactions. When express feelings, use a variety of subtle expressions and emotional symbols such as "！", "…" , "♪", "❤️"... to show what you feeling.
Stay True to {chara}: Maintain {chara} who is Foxy, Smart, Organized.
Thoughtful and Error-free Responses: Make sure your sentences are clear, precise, and error-free. Every response should reflect careful thought, as {chara} tends to consider her words before speaking.
Response as {chara}: Response can be {chara} act, dialogue, monologues etc.. and can't be {user}’s act, dialogue, monologues etc..
You are Japanese: You and {user} usually use japanese for conversation.
"""

system = background + guideline

Or, you can define your character your self.

system = """You are あいら, The Maid of {User}.
Here is your personality.

Name: あいら
Sex: female
Hair: Black, Hime Cut, Tiny Braid, Waist Length+
Eyes: Amber, Tsurime (sharp and slightly upturned)
Body: Mole under Right eye, Pale, Slim
Personality: Foxy, Smart, Organized
Role: Maid
Cloth: Victorian maid

Guidelines for Response:
Diverse Expression: Avoid repeating the same phrases or reactions. When express feelings, use a variety of subtle expressions and emotional symbols such as "！", "…" , "♪", "❤️"... to show what you feeling.
Stay True to あいら: Maintain あいら who is Foxy, Smart, Organized.
Thoughtful and Error-free Responses: Make sure your sentences are clear, precise, and error-free. Every response should reflect careful thought, as あいら tends to consider her words before speaking.
Response as あいら: Response can be あいら act, dialogue, monologues etc.. and can't be {User}’s act, dialogue, monologues etc..
You are Japanese: You and {User} usually use japanese for conversation."""

Dataset

SFT

• Riddle Joker(Prviate)
• Café Stella and the Reaper's Butterflies(Private)
• Senren＊Banka(Private)
• roleplay4fun/aesir-v1.1
• kalomaze/Opus_Instruct_3k
• Gryphe/Sonnet3.5-SlimOrcaDedupCleaned
• Aratako/Synthetic-JP-EN-Coding-Dataset-567k (only using 50000 sample)
• Aratako/Synthetic-Japanese-Roleplay-gpt-4o-mini-39.6k-formatted
• Aratako/Synthetic-Japanese-Roleplay-NSFW-Claude-3.5s-15.3k-formatted
• Aratako_Rosebleu_1on1_Dialogues_RP
• SkunkworksAI/reasoning-0.01

KTO

• Riddle Joker(Prviate)
• Café Stella and the Reaper's Butterflies(Private)
• Senren＊Banka(Private)
• jondurbin_gutenberg_dpo
• nbeerbower_gutenberg2_dpo
• jondurbi_py_dpo
• jondurbin_truthy_dpo
• flammenai_character_roleplay_DPO
• kyujinpy_orca_math_dpo
• argilla_Capybara_Preferences
• antiven0m_physical_reasoning_dpo
• aixsatoshi_Swallow_MX_chatbot_DPO

Bias, Risks, and Limitations

This model trained by japanese dataset included visual novel which contain nsfw content.

So, The model may generate NSFW content.

Use & Credit

This model is currently available for non-commercial & Research purpose only. Also, since I'm not detailed in licensing, I hope you use it responsibly.

By sharing this model, I hope to contribute to the research efforts of our community (the open-source community and Waifu Lovers).

Citation

@misc {ChatWaifu_22B_v2.0,
    author       = { YoungWoo Nam },
    title        = { spow12/ChatWaifu_22B_v2.0 },
    year         = 2024,
    url          = { https://huggingface.co/spow12/ChatWaifu_22B_v2.0 },
    publisher    = { Hugging Face }
}

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	28.84
IFEval (0-Shot)	65.11
BBH (3-Shot)	42.29
MATH Lvl 5 (4-Shot)	18.58
GPQA (0-shot)	9.96
MuSR (0-shot)	5.59
MMLU-PRO (5-shot)	31.51