Multi-Layer XAI Framework for Log Anomaly Detection - Processed Artifacts (BGL)

Processed artifacts for the multi-layer XAI framework built on top of LAnoBERT (bert-base-uncased fine-tuned on normal BGL log lines using masked language modeling). Contents: - BERT embeddings for 100,000 normal and 100,000 anomaly BGL log lines (float32, 768-dim) - Train/val/test split CSVs (70/10/20 split, seeded at 42) - Anomaly threshold and scoring metrics (threshold_bgl.json) - Fine-tuning loss history and summary - Pre-computed t-SNE embedding plot and XAI output figures Fine-tuned model weights are hosted separately on HuggingFace: https://huggingface.co/Buggycat/lanobert-bgl Source code: https://github.com/buggicat/INSURE_2026 Dataset: BGL (Blue Gene/L supercomputer logs) from LogHub — https://github.com/logpai/loghub