This repository includes the training codes for SuperLog, as well as a specifically designed interpretable dataset for the log analysis domain, named NLPLog.
NLPLog is a question-answer dataset generated through interactions with ChatGPT. We utilized the publicly available log datasets provided by LogHub, posing questions to ChatGPT across five defined dimensions. This approach yielded conversational data rich in log-related domain information and high interpretability. NLPLog will serve as the dataset for the subsequent continuous pre-training of SuperLog.
NLPLog data can be get here.
At this stage, we leverage the open-source LLM training framework, LLaMa-Factory, to continuously pre-train an open-source large model using the NLPLog dataset. In this step, we infuse interpretable knowledge, equipping the general-purpose LLM with domain-specific information related to log analysis.
- Environment Configuration
pip install -e ".[torch,metrics,deepspeed]"- Start Training
llamafactory-cli train examples/train_full/train_superlog.yamlIn this step, we fine-tune SuperLog using a selection of 1000 high-quality instruction data from Alpacar. The purpose of this step is to enable SuperLog to have a good ability to respond to human instructions, understand log analysis commands, and complete the corresponding tasks.
llamafactory-cli train examples/train_full/iftrain_superlog.yaml