MarkrAI/RAG-KO-Mixtral-7Bx2-v2.0
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: 5
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-v2.0"
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...