Optimizing Llama-3-8B for cultural reasoning for a 2 month project at TU Dresden.
Standard Large Language Models often exhibit Western-centric biases and lack the nuance required for specific cultural queries. Additionally, the 8B parameter model struggled with strict output formatting (JSON) and zero-shot reasoning for complex Short Answer Questions (SAQ), often leading to "instruction drift".
I stripped away complex JSON prompts in favor of natural language instructions and built a "Dynamic Few-Shot" RAG framework. By transforming multiple-choice datasets into direct QA pairs, the system dynamically retrieves and injects the top 3 most relevant examples into every prompt.
The dual approach of LoRA fine-tuning and Dynamic RAG crushed the baseline zero-shot performance. The system outputted logically sound, culturally accurate answers required by the strict academic evaluator, resulting in a definitive 16% accuracy gain.
The solution utilizes the Llama-3-8B model with 4-bit quantization (NF4) for efficiency. It employs `all-MiniLM-L6-v2` for semantic embedding and retrieval. The pipeline includes a data augmentation stage where training data is stripped of options to create direct QA pairs. Inference uses Greedy Search to ensure deterministic and concise outputs required by the evaluation script.
Retrieval system that injects semantically relevant "in-context" examples for each specific query, boosting SAQ accuracy by 16%.
Automated pipeline to transform MCQ datasets into SAQ pairs, effectively doubling the training resources for the retrieval corpus.
Used LoRA (Low-Rank Adaptation) and quantization to fine-tune the 8B model on limited hardware, optimizing for task alignment.
Evaluated external internet search (DuckDuckGo), discovering that "clean" internal data outperforms noisy web results for this specific domain.
Setting up the environment with Transformers/PEFT and executing initial Fine-Tuning runs.
Developing the Dynamic Few-Shot logic and constructing the augmented knowledge base.
Running ablation studies on decoding strategies and external search integration.
Finalizing the project report and analysis of the 0.16 accuracy gain.