LightChatAssistant-4x7B
GGUF版はこちら/Click here for the GGUF version
概要
Sdff-Ltba/LightChatAssistant-2x7Bと全く同じ手法を用いつつ、対象を4モデルに拡張してそれらをMoEしたモデルです。素晴らしい手法を提案いただいた@Sdff-Ltbaさんに感謝します。
モデルサイズは大きくなっていますが、推論に同時に使われるエキスパートは2つなので、GPUにオフロードさえできれば元のSdff-Ltba/LightChatAssistant-2x7Bとほぼ同じような速度で動きます。
以下の4モデルをChat Vectorの加算によって対話強化して素材に使用しています。
- NTQAI/chatntq-ja-7b-v1.0
- stabilityai/japanese-stablelm-instruct-gamma-7b
- Elizezen/Antler-7B
- Elizezen/Hameln-japanese-mistral-7B
Chat Vectorの適用方法はSdff-Ltba/LightChatAssistant-2x7Bと同じで、Mistral-7B-Instruct-v0.2とMistral-7B-v0.1の差分を0.8倍してそれぞれのモデルに足しています。
その後、mergekitを使い、4モデルをMoEしました。MoEのconfigは以下の通りです。
base_model: ./chatntq-chatvector
gate_mode: hidden
dtype: bfloat16
experts:
- source_model: ./chatntq-chatvector
positive_prompts:
- "roleplay"
- "question"
- "answer"
- "chat"
- "companion"
- "character"
- "math"
- "code"
- "[Mode: Roleplay]"
- "[Mode: Chat]"
negative_prompts:
- "storywriting"
- "book"
- "story"
- "chapter"
- "tale"
- "history"
- "write"
- "novel"
- "[Mode: Writing]"
- source_model: ./japanese-stablelm-instruct-gamma-chatvector
positive_prompts:
- "roleplay"
- "question"
- "answer"
- "chat"
- "companion"
- "character"
- "math"
- "code"
- "[Mode: Roleplay]"
- "[Mode: Chat]"
negative_prompts:
- "storywriting"
- "book"
- "story"
- "chapter"
- "tale"
- "history"
- "write"
- "novel"
- "[Mode: Writing]"
- source_model: ./Hameln-japanese-mistral-chatvector
positive_prompts:
- "sex"
- "storywriting"
- "erotic"
- "fuck"
- "orgasm"
- "uncensored"
- "book"
- "story"
- "chapter"
- "tale"
- "history"
- "write"
- "novel"
- "[Mode: Writing]"
- source_model: ./Antler-chatvector
positive_prompts:
- "sex"
- "storywriting"
- "erotic"
- "fuck"
- "orgasm"
- "uncensored"
- "book"
- "story"
- "chapter"
- "tale"
- "history"
- "write"
- "novel"
- "[Mode: Writing]"
tokenizer_source: union
- Downloads last month
- 39
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social
visibility and check back later, or deploy to Inference Endpoints (dedicated)
instead.