Higher-bit quantisation (8-bit) better preserves structural coherence and semantic alignment in a Pidgin-fine-tuned LLM, achieving the highest ROUGE-L (0.3120) and BERTScore (0.8590) among the quantised models¶

Claim · All claims
Evidence · 1 study · design r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? Among the fine-tuned quantised models, LLM-Q8 achieved the highest ROUGE-L (0.3120) and BERTScore (0.8590), indicating better structural coherence and semantic preservation than lower-bit models. → Nwogo 2026

Evidence¶

Nwogo 2026¶

Nwogo, E. U., Ihianle, I. K., Machado, P., Bird, J. J., Lotfi, A., Shuaib, A. A., Akinwumi, I. I., & Oluranti, J. (2026). An AI-Based Adaptive Learning Platform for Multilingual and Low-Resource Educational Contexts: A Case Study on Nigeria. arXiv. https://arxiv.org/abs/2608.15738 doi:10.48550/arxiv.2608.15738

q2 · i? · design · r2

Automatic semantic evaluation of three fine-tuned, quantised LLMs (Q4, Q5, Q8) on 14 science prompts with gpt-4o outputs as references. LLM-Q8 "achieved the highest ROUGE-L score (0.3120)" and the highest BERTScore (0.8590) in Table 1; BLEU was uniformly low (0.0120–0.0121).

"However, higher-bit quantisation (LLM-Q8) achieved the highest ROUGE-L score (0.3120), suggesting that in- creased numerical precision better preserves structural co- herence and sentence-level flow."

Discussion¶

Related Claims¶

• The OpenAI gpt-4o reference model outperformed all quantised models on human-rated coherence, fluency, relevance and cultural appropriateness, serving as an upper benchmark — related
• Native speakers rated outputs of LLMs fine-tuned on Nigerian Pidgin English as understandable and context-aware, though no non-fine-tuned baseline was evaluated — related
• A 4-bit quantized 27B local Teacher LLM generated knowledge in the required format, suggesting accessible-resource implementations may be feasible — related
• Cluster membership showed a strong multivariate effect on standard text-similarity metrics (Pillai's trace = 0.85, F(12, 248) = 15.31, p < 0.001) — related
• Hallucination rate increases under 4-bit quantization and is substantially mitigated by quantization-aware fine-tuning — related