log-classifier-BERT-v1

log-classifier-BERT-v1

This model is a transformers classification model, trained using BERTForSequenceClassification designed for use in network and device log mining tasks. Developed by Selector AI

Model Usage

from transformers import BertForSequenceClassification, BertTokenizer

# Step 1: Load the model and tokenizer from Hugging Face
model = BertForSequenceClassification.from_pretrained("rahulm-selector/log-classifier-BERT-v1")
tokenizer = BertTokenizer.from_pretrained("rahulm-selector/log-classifier-BERT-v1")

import torch

model.eval()

# Step 2: Prepare the input data (Example log text)
log_text = "Error occurred while accessing the database."

# Tokenize the input data
inputs = tokenizer(log_text, return_tensors="pt", padding=True, truncation=True, max_length=128)

# Step 3: Make predictions
with torch.no_grad():
    outputs = model(**inputs)
    logits = outputs.logits

# Step 4: Get the predicted class (the class with the highest score)
predicted_class = torch.argmax(logits, dim=1).item()

# label mapping (can load from JSON file in repo or config)
label_mapping = model.config.id2label

# Step 5: Get the event name
predicted_event = label_mapping[predicted_class]
print(f"Predicted Event: {predicted_event}")

Background

The model focuses on structured and semi-structured log data, outputing around 60 different event categories. It is highly effective for real-time log analysis, anomaly detection, and operational monitoring, helping organizations manage large-scale network data by automatically classifying logs into predefined categories, facilitating faster and more accurate diagnosis of network issues.

Intended uses

Our model is intended to be used as classifier. Given an input text (a log coming from a network/device/router), it outputs a corresponding event most associated with the log. The possible events that can be classified are shown in encoder-main.json

Training Details

Data

The model was trained on a variety of network events and system logs, focusing on monitoring and analyzing state changes, protocol behaviors, and hardware interactions across infrastructure components. This included tracking routing issues, protocol neighbor state changes, link stability, and security events, ensuring that the model could recognize and classify critical patterns in device communications, network health, and configuration activities.

Train/Test Split

• Train Data Size: ~80K Logs
• Test Data Size: ~20K Logs

Hyper Parameters

The following hyperparameters were used during training to optimize the model's performance:

• Batch Size: 32
• Learning Rate: .001
• Optimizer: Adam
• Epochs: 10
• Dropout Rate: N/A
• LSTM Hidden Dimension: 384
• Embedding Dimension: 384