RAG-KO-Mixtral-7Bx2-v2.1

MarkrAI/RAG-KO-Mixtral-7Bx2-v2.1

Model Details

Model Developers

MarkrAI - AI Researchers

Base Model

DopeorNope/Ko-Mixtral-v1.4-MoE-7Bx2.

Instruction tuning Method

Using QLoRA.

4-bit quantization
Lora_r: 64
Lora_alpha: 64
Lora_dropout: 0.05
Lora_target_modules: [embed_tokens, q_proj, k_proj, v_proj, o_proj, gate, w1, w2, w3, lm_head]

Hyperparameters

Epoch: 10
Batch size: 64
Learning_rate: 1e-5
Learning scheduler: linear
Warmup_ratio: 0.06

Datasets

Private datasets: HumanF-MarkrAI/Korean-RAG-ver2

Aihub datasets 활용하여서 제작함.  

Implmentation Code

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

repo = "MarkrAI/RAG-KO-Mixtral-7Bx2-v2.1"
markrAI_RAG = AutoModelForCausalLM.from_pretrained(
        repo,
        return_dict=True,
        torch_dtype=torch.float16,
        device_map='auto'
)
markrAI_RAG_tokenizer = AutoTokenizer.from_pretrained(repo)

Model Benchmark

• Coming soon...