Benchmarks - LogLens AI · F1 0.957, 1.000 recall on Loghub BGL
// Loghub BGL · 500,000 lines · 206,847 labeled alerts

Numbers you can run yourself.

Every stat below comes from public labeled datasets. The benchmark command is in the box at the bottom - reproduce any of it on your own machine.

Headline accuracy - Loghub BGL

All three modes catch every one of the 206,847 alerts. Deep mode trims false positives further.

Mode Engine Precision Recall F1 Speed Missed
fast from-scratch statistical 0.901 1.000 0.948 ~6,700/s 0
turbo optimized statistical 0.901 1.000 0.948 ~7,300/s 0
deep AI semantic embeddings 0.917 1.000 0.957 ~3,400/s 0

Precision · Recall · F1 by mode

Precision Recall F1 axis 0.85-1.00
{{ barGroups }}

Cross-dataset generality

Thunderbird - 500,000 all-normal lines, no retuning. A near-perfect specificity means almost no false alarms.

{{ gauges }}
CAPABILITY · NOT A BENCHMARK

15+ log formats - detected automatically, zero config.

One detector, every stack - no parsers to configure, no schema to define. Point it at a file and it figures out the format.

☁️Cloud & structuredenterprise
AWS (CloudTrail) GCP (Cloud Logging) Azure (Monitor/Activity) JSON
🖥️Infra & web
NginxApache accessApache errorSyslogHPC
🧱Big-data & platform
HDFSSparkZookeeper
📦App & endpoint
Java / AppHealthAppWindows CBSProxifier

…plus a generic fallback that handles anything else - timestamps, levels, and messages are extracted even from formats it hasn't seen.

0/30
injected incidents caught

100% incident recall across 6 real datasets.

30 injected critical incidents - kernel panic, OOM, disk failure, security breach, data corruption - hidden in normal traffic. LogLens caught every one, zero configuration. These are Loghub evaluation datasets used to measure detection accuracy - distinct from the auto-detected formats above.

Apache
✓ 5/5
Spark
✓ 5/5
HDFS
✓ 5/5
HealthApp
✓ 5/5
OpenStack
✓ 5/5
Thunderbird
✓ 5/5

Throughput

Statistical modes stream through hundreds of thousands of lines per minute. On the 810-line demo, loglens bench hits 14,999 lines/s and collapses them into 8 incident families.

0 lines/s · turbo bench
Lines per second by mode
6,700
fast
7,300
turbo
3,400
deep
REPRODUCIBLE

Don't trust the numbers. Run them.

Point the benchmark at any line-labeled log and it prints precision, recall, and F1. A CI gate fails the build if F1 drops below your threshold.

$ loglens benchmark labeled.log --min-f1 0.90
⚖️

Honesty notes

·

Accuracy is measured against Loghub's line-level labels - the same public ground truth other tools are evaluated on.

·

Deep mode embeds unique templates only, not every line. This is disclosed and is why it stays fast on huge files.

·

RCA and ask send only grouped summaries to the LLM - never your full logs. Detection itself is always local.

·

Alerting works fully offline. The optional LLM layer only enriches an alert that would have fired anyway.