| base_model: Lalaa-a/gemma4-8b-text-only | |
| library_name: peft | |
| license: gemma | |
| tags: [sinhala, mmlu, multiple-choice, lora] | |
| # Sinhala MMLU: LoRA adapter + choice head | |
| LoRA adapter (rank 32, all 24 kept layers) for | |
| [Lalaa-a/gemma4-8b-text-only](https://huggingface.co/Lalaa-a/gemma4-8b-text-only) (text weights of | |
| google/gemma-4-E4B-it) cut after layer 24, plus `head.pt`: a small transformer choice head that scores each | |
| option from Gemma's hidden states. Not usable with plain `AutoModelForCausalLM`: load it with the team's | |
| `inference.py` (main repository), which rebuilds the cut model from `meta.json`. | |
| | File | What | | |
| |---|---| | |
| | `lora/` | PEFT adapter (`adapter_config.json`, `adapter_model.safetensors`) | | |
| | `head.pt` | choice head weights (PyTorch state dict) | | |
| | `meta.json` | how to rebuild the network: layer cut, head size, token limits, readout | | |
| | `config.yaml` | the full training config of this checkpoint | | |