Task-Specific Full Fine-Tuning for Retrieval-Augmented Generation: A Multi-Family Evaluation Across Open-Weight Language Models

Retrieval-Augmented Generation (RAG) depends not only on retrieval quality but alsoon the generator’s ability to use supplied evidence, abstain when support is insufficient, andfollow task-specific output requirements. While prior work has shown that such behaviourscan be improved through targeted adaptation, less is known about how the same RAG-specifictraining corpus affects different model families and parameter scales. We conduct a controlledempirical study of task-specific full fine-tuning across ten open-weight model configurationsfrom the Gemma, Llama, and Qwen families, ranging from approximately 0.5B to 14Bparameters. All models are exposed to the same fixed German-language training corpus andevaluated in a paired base-versus-fine-tuned design on identical test instances. Fine-tuningimproves the aggregate benchmark score for all ten configurations, but the magnitude anddistribution of gains vary substantially. The strongest and most consistent improvementoccurs in hard-negative handling, whereas grounded question-answering scores decrease fornine of ten fine-tuned models under the benchmark metric. Structured output and citationbehaviour generally improve, but not uniformly. Within the Qwen family, model size isassociated with stronger absolute performance but not with larger fine-tuning gains, andseveral smaller fine-tuned models outperform larger untuned models from the same family onthe evaluated task distribution. These results support interpreting task-specific full fine-tuningas behavioural adaptation rather than uniform capability enhancement. Model size, baseperformance, final task performance, and adaptation gain should therefore be treated asdistinct considerations when selecting and adapting generators for specialized RAG systems.