---
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.