Early-Layer vs. Late-Layer LoRA Fine-Tuning for Cross-Lingual NLI in Low-Resource African Languages
Domain:
natural language processingRecord type:
paper
Creator:
Ass
Publisher:
Host:
Large Language Models (LLMs) have demonstrated remarkable capabilities, yet their performance in low-resource languages (LRLs), such as Swahili, often lags due to data scarcity and underrepresentation in pre-training. A key challenge is achieving robust cross-lingual lexical alignment, crucial for tasks like translation and cross-lingual information retrieval. This paper introduces Targeted Lexical Injection (TLI), a novel and efficient fine-tuning approach. We first demonstrate that Lugha-Llama-8B-wura, a Swahili-centric LLM, exhibits strong, near-perfect lexical alignment for Swahili-English
Research goal: How does early-layer LoRA fine-tuning with TLI compare to late-layer LoRA fine-tuning in terms of cross-lingual natural language inference accuracy on the XNLI benchmark for severely low-resource African languages?
Autonomous synthesis report generated by Assignee Research. Tribunal consensus score: 7.5/10. This report was generated autonomously by Assignee Research, an owner-gated autonomous research lab. The content synthesizes findings from peer-reviewed papers. Tribunal score: 7.5/10.
Similar
Early-Layer LoRA Fine-Tuning for Zero-Shot Cross-Lingual NLI in Low-Resource African LanguagesEarly-Layer LoRA Fine-Tuning Robustness in Cross-Lingual NLI for Underrepresented LanguagesEarly-Layer LoRA Fine-Tuning vs Full-Parameter Tuning for Zero-Shot Cross-Lingual Retrieval in Low-Resource LanguagesEarly-Layer LoRA Adaptation vs Full Fine-Tuning for Zero-Shot Cross-Lingual Transfer in Low-Resource African LanguagesLoRA Rank Variation in Early-Layer Fine-Tuning and Cross-Lingual Alignment for Low-Resource African LanguagesLoRA Rank Variation in Middle-Layer vs. Late-Layer Fine-Tuning for Swahili-English Cross-Lingual Alignment