team-qbits-cpt-v3
Submitted system for the ICTer 2026 SinhalaMMLU Shared Task (macro 57.81).
Base: QBitsLK/gemma4-e4b-sinhala-cpt (Gemma 4 E4B + Sinhala continued pretraining). LoRA r=32, alpha=32, lr 1e-4, 2 epochs on 21,235 past-paper MCQ examples; checkpoint final_adapter (step 1328), merged. 7,996,157,418 parameters.
Inference: per-option log-likelihood under a 5-prompt ensemble. See SinhalaMMLU-SharedTask/team-qbits.
- Downloads last month
- 27
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for SinhalaMMLU-SharedTask/team-qbits-cpt-v3
Base model
QBitsLK/gemma4-e4b-sinhala-cpt