metadata
license: mit
library_name: pytorch
tags:
- gpt2
- causal-lm
- log-analysis
- cybersecurity
- devops
metrics:
- perplexity
- rougeL
- bertscore
LoGPT — GPT-2 Fine-Tuned for Log Analysis
LoGPT is a GPT-2 (124M) causal language model fine-tuned for structured system log analysis.
It generates:
- Log summaries
- Root cause analysis
- Actionable recommendations
Runs fully offline on local hardware.
Model Details
- Base model: GPT-2 (124M)
- Architecture: Decoder-only Transformer
- Context length: 1,024 tokens
- Training samples: 5,055
- Fine-tuning steps: 2,000
Evaluation (Test Set)
- Perplexity: 2.05
- ROUGE-L: 0.324
- BERTScore F1: 0.898
Intended Use
Designed for structured system/server logs (Apache, Linux, HDFS, Spark, OpenSSH, etc.).
Not intended for:
- General chat
- Medical / legal advice
- Safety-critical automation
Prompt Format
<|log_start|> log content <|log_end|> <|query_start|> Summarize these logs <|query_end|> <|response_start|>
Model generates until <|response_end|>.
Limitations
- Trained on ~5k synthetic samples
- May hallucinate plausible root causes
- Performance may drop on unseen log formats
Human verification recommended.