In this paper we propose self-evolving method called SelfLog，which, on one hand, uses similar <group, template> pairs extracted by LLM itself in the historical data to act as the prompt of a new log, allowing the model to learn in a self-evolution and labeling-free way. On the other hand, we propose an N-Gram-based grouper and log hitter.
├── evaluate/ #
│ ├── evaluator/ # the evaluation code of GA, PA, PTA, RTA
│ └── evaluator_PA/ # calculate PA, PTA, RTA result
├── functions/ # mian part of SelfLog
│ ├── benchmark_settings/ # log data process
│ ├── gram/ # N-gram based grouper
│ ├── llm_func/ # requst llm
│ └── tree_based_merge/ # the postprocess of SelfLog
├── logs/
│ └── ...... # parsing log files
├── online_selfLog/ # online version of SelfLog
│ ├── is_new_log # log hitter
│ ├── log_pruduce # streaming log production
│ └── online_run # test the efficient of SelfLog
├── PSQL/ # Prompt database recall method based on PostgreSQL
│ ├── model # the embedding model of SelfLog
│ ├── conConfig # connect psql setting
│ ├── exampleToPSQL # algorithm startup candidate set written to psql
│ └── findTopKexam # recall examples
├── CONSTANT # hyperparameter configuration items
├── llmAPIsetting # llm address url and key
├── prompt # llm prompt format
├── run.py # test the effect of SelfLog on the dataset
└── README.md
- Prompt Database We use psql with the vector plugin to implement a method for retrieving and recalling related logs based on semantic similarity. You can also use other databases for your purposes.
- Install PostgresSQL
- Creat table
such as
CREATE TABLE IF NOT EXISTS public.log_template
(
"ID" integer NOT NULL DEFAULT nextval('id_seq'::regclass),
log text COLLATE pg_catalog."default",
template text COLLATE pg_catalog."default",
"logVector" vector,
CONSTRAINT seflog_pkey PRIMARY KEY ("ID")
);
- Python
- Install python >= 3.8
- pip install -r requirements.txt
- LLM API
- API-key
- model url
- Candidates to prompt database
- cd PSQL
- python exampleToPSQL.py
- Effect evaluation
- python run.py
The analysis results will be stored in the log directory.
- Efficiency evaluation
- cd online_selfLog
- download full dataset
- python log_pruduce.py