README.md

---
license: gemma
base_model: unsloth/gemma-4-E2B-it
tags:
  - lora
  - classification
  - telelogs
  - 5g-rca
---

# TeleLogs 5G Root Cause Classifier (Gemma-4-E2B + LoRA)

LoRA fine-tune of unsloth/gemma-4-E2B-it (language layers only) with a custom
classification head (Dropout -> LayerNorm -> Linear) for 8-way root cause
classification (C1-C8) on the netop/TeleLogs dataset.

## Files
- adapter_model.safetensors / adapter_config.json: LoRA adapter
- layer_norm.pt: LayerNorm state dict applied to final hidden state
- classifier_head.pt: Linear classifier head state dict
- classification_config.json: hidden_size, num_labels, id2label mapping

## Usage
See the accompanying `app.py` in the linked Space for a full CPU-compatible
loading example (base model + LoRA adapter + custom head, loaded with plain
transformers + peft rather than Unsloth).

Accuracy: 0.9444 | Macro F1: 0.9444