DevQuestKhuselo/RESEARCH

This research investigates whether post-training quantisation (GPTQ) and magnitude-based pruning disproportionately degrade multilingual translation quality for low-resource African language pairs compared to high-resource language pairs in large language models. # Evaluating the Impact of Quantisation and Pruning on Multilingual LLM Translation for Low-Resource African Languages **Author:** Khuselo Sofohlo **Supervisors:** Dr Steven James, Prof Benjamin Rosman **Institution:** School of Computer Science and Applied Mathematics, University of the Witwatersrand **Course:** COMS4057A Research Project ## Overview This research investigates whether post-training quantisation (GPTQ) and magnitude-based pruning disproportionately degrade multilingual translation quality for low-resource African language pairs compared to high-resource language pairs in large language models. Model compression is essential for efficient LLM deployment, but most compression research evaluates degradation on English and high-resource benchmarks, implicitly assuming uniform degradation across languages. This assumption may not hold for low-resource languages with weaker cross-lingual representations. This project systematically evaluates that assumption using empirical translation-quality measurements before and after compression. ## Research Question To what extent do quantisation and pruning disproportionately affect translation quality for low-resource African language pairs compared to high-resource language pairs in multilingual large language models? ## Hypothesis Low-resource translation directions will experience significantly larger degradation in translation quality after quantisation and pruning than high-resource translation directions, due to weaker and less robust multilingual representations. ## Model **LLaMAX3-8B-Alpaca** — an instruction-tuned, decoder-only LLM built on continued multilingual pretraining of Llama-3-8B, supporting 100+ languages including Swahili and isiZulu. > **Note on model selection:** the original proposal specified Aya-8B (Aya Expanse). During setup it was confirmed that Aya Expanse 8B only supports 23 languages, none of which are isiZulu or Swahili, making it incompatible with this study's translation …