Team Nena – Sinhala MMLU Shared Task System

System summary

Base model(s): SAWithanage/SinLlama-Llama-3-8B-Merged
What we did: We performed parameter-efficient fine-tuning (LoRA) of SinLlama-8B on authentic Sri Lankan national curriculum examination Multiple-Choice Questions (spanning Grades 6–13 across 14 official subjects) alongside domain-adaptive curriculum pretraining passages. For inference, we employ a Sinhala Micro-Chain-of-Thought (Micro-CoT) reasoning prompt with deterministic greedy decoding to ground curriculum facts before emitting the final anchored choice.
Total inference parameters: 7.954B parameters (Language-based vocabulary trimming (Lang-Trim) slices out 20,336 inactive foreign script tokens while preserving 100% of Sinhala, English, numbers, and symbols, safely conforming to the Hard Rule 02 parameter ceiling of 8.00B as officially approved by the shared task chairs).
Extra data used for training or retrieval: Curated Sri Lankan national curriculum examination MCQs and government textbook passages across Grades 6–13. Strictly retrieval-free at inference time (zero external APIs, web search, or RAG).

Repo	What it is
`SinhalaMMLU-SharedTask/team-nena`	Code, inference script, final predictions, and training scripts (this repo)
`SinhalaMMLU-SharedTask/team-nena-sinllama-8b-lora`	Fine-tuned LoRA adapter weights and tokenizer configuration

pip install -r requirements.txt
python inference.py

GPU used: 1x Nvidia A100 (80GB / 40GB) or 1x T4 / RTX 3090 / 4090 (with 4-bit NF4)
Approximate run time on the test set: ~80–90 minutes for 2,238 test questions
Anything else the organizers need to know: Pure offline inference, zero closed or paid APIs.

inference.py – the exact script that produced our final Codabench submission
predictions.jsonl – the exact file we submitted to the final leaderboard (2,238 predictions)
requirements.txt – package versions used
training/ – training, fine-tuning, and data-preparation code and configurations