config.yaml

# is_train: True # train deepsvdd
# is_rag: True # use post processing stage
# dataset_name: BGL
# train_ratio: 0.8
# window_size: 200
# window_time: 1800 # 30 minutes

# # dataset and model paths
# log_structed_path: ./dataset/BGL/bgl-example.log_structured.csv
# encoder_path: ./pretrained

# # openai key
# api_key: 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
# api_base: https://api.openai-proxy.org/v1

# llm_name: gpt-3.5-turbo
# # llm_name: mistralai/Mistral-7B-Instruct-v0.1 # from huggingface

# # deepsvdd parameters
# normal_class: 0
# is_pretrain: False
# optimizer_name: adam
# lr: 0.0001
# n_epochs: 150
# lr_milestones: [50]
# batch_size: 40960
# weight_decay: 0.0005
# device: cuda:0
# n_jobs_dataloader: 0

# # rag parameters
# threshold: 0.8
# topk: 5
# prompt: prompt5
# persist_directory: ./output/ragdb-bgl

# LogRAG Configuration (Mistral-7B Version)
# =========================================

# Execution Mode
is_train: True # Whether to train DeepSVDD
is_rag: True # Whether to use RAG post-processing
dataset_name: BGL # Dataset to use

# Dataset Parameters
train_ratio: 0.8 # Train/test split ratio
window_size: 200 # Number of logs per window
window_time: 1800 # Time window in seconds (30 mins)

# Model Configuration
llm_name: "mistralai/Mistral-7B-Instruct-v0.1" # Mistral model
encoder_path: "./pretrained/mistral" # Local model cache

# Dataset Paths
log_structed_path: "./dataset/BGL/bgl-example.log_structured.csv"

# DeepSVDD Training Parameters (Adjusted for Mistral)
normal_class: 0 # Class treated as normal
is_pretrain: False # Whether to pretrain
optimizer_name: adam # Optimization algorithm
lr: 0.0001 # Learning rate (may need adjustment)
n_epochs: 50 # Reduced epochs for efficiency
lr_milestones: [20] # Adjusted milestones
batch_size: 8 # Significantly reduced for VRAM constraints
weight_decay: 0.0005 # L2 regularization
device: "cuda:0" # Requires GPU with 24GB+ VRAM
n_jobs_dataloader: 0 # Workers for data loading

# RAG Parameters (Mistral-optimized)
threshold: 0.7 # Adjusted similarity threshold
topk: 3 # Reduced retrieved passages
prompt: "mistral_prompt" # Special prompt template for Mistral
persist_directory: "./output/ragdb-bgl-mistral" # Separate vector store

# Quantization Configuration (Recommended)
use_4bit: True # Enable 4-bit quantization
bnb_4bit_type: "nf4" # Normal Float 4 quantization