MarkrAI/Ko-mistral-7B-Inst-Wizard-v2.0-epoch3

Base Model: mistralai/Mistral-7B-Instruct-v0.2.

Training Datasets

HQ-Korea-Datasets: WizardLM_Evol_train.
Learning rate: 4e-6

Validation Datasets

HQ-Korea-Valid: WizardLM_Evol_valid.
Eval loss: 0.4612

Training Method.

The following papers contain the foundational methodologies for the dataset and training methods we are currently proceeding.

• LIMA.
• WizardLM.
• Near Dedup.

MT-Bench score

• LogicKor.

  Model	추론	수학	글쓰기	코딩	이해	문법	싱글턴	멀티턴	평균
  claude-3-opus-20240229	8.42	9.21	9.71	9.14	10.00	7.92	8.80	9.33	9.07
  gpt-4-turbo-2024-04-09	9.07	9.85	9.78	9.50	9.14	6.42	9.07	8.85	8.96
  HyperClovaX	5.85	7.14	8.50	7.57	9.50	8.50	8.40	7.28	7.84
  maywell_kiqu-70b	7.35	6.14	8.92	7.85	8.28	5.71	8.16	6.59	7.38
  google-gemini-1.5-pro	6.50	6.92	7.78	8.28	7.78	5.21	7.90	6.26	7.08
  solar-1-mini-chat	6.35	4.28	8.50	6.71	7.00	5.21	6.42	6.26	6.34
  mistralai_Mixtral-8x7B-Instruct-v0.1	5.35	4.21	5.42	5.64	6.42	3.21	5.42	4.66	5.04
  MarkrAI/Ko-mistral-7B-Inst-Wizard-v2.0-epoch3 (ours)	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN

자체 GPT4 평가(싱글턴; 3번 평균): 7.4682