SinhalaMMLU-SharedTask/team-nena-sinllama-8b-lora

Team Nena: SinLlama-8B LoRA Adapter for SinhalaMMLU Shared Task (ICTer 2026)

This repository contains the official fine-tuned LoRA adapter developed by Team Nena for the SinhalaMMLU Shared Task at ICTer 2026.

Model Overview

• Base Model: SAWithanage/SinLlama-Llama-3-8B-Merged
• Adapter Type: LoRA (Rank = 32, Alpha = 64)
• Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Target Task: Standalone, offline Sinhala Multiple-Choice Question Answering across the Sri Lankan national curriculum (Grades 6–13 across all 14 official subjects).
• Inference Mode: Sinhala Micro-Chain-of-Thought (Micro-CoT) generative reasoning with deterministic greedy decoding.

Rule 02 Parameter Compliance

• Base Backbone: 32 Transformer layers intact.
• Vocabulary Slicing: Foreign script subwords pruned from 139k to 119,000 via language-based vocabulary trimming (Lang-Trim).
• Total Inference Parameters: 7,954,436,096 parameters (7.954B), which is 45.56 Million safely below the 8.00B parameter limit, officially approved by the shared task chairs.

Main System Repository

The inference code, data preparation scripts, and final submission predictions are available in the main team repository:

• SinhalaMMLU-SharedTask/team-nena