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
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support