L3D-RAG: Leveraging LLaMA 3.1 and DeepSeek with Retrieval-Augmented Generation (RAG) for advanced Reddit community analysis

This study introduces L3D-RAG, a generative framework designed to evaluate the performance of two generative AI models: LLaMA 3.1 and DeepSeek. Each model is integrated independently with a retrieval-augmented generation (RAG) mechanism for analyzing Reddit discourse, enabling accurate and context-aware modeling of Reddit, as a large-scale and thematically diverse platform, provides a rich resource for studying public discourse and community trends; however, its linguistic diversity and structural complexity challenge conventional language models, making precise classification vital for content understanding and moderation. LLaMA 3.1 and DeepSeek were fine-tuned with Low-Rank Adaptation (LoRA) and 4-bit quantization for memory-efficient deployment. In L3D-RAG, LLaMA 3.1 serves as a context-aware generator and DeepSeek as a semantic retriever, enabling direct comparison of generation- and retrieval-focused configurations on Reddit classification and calibration tasks. Experimental results show that LLaMA 3.1 + RAG achieved an average F1-score of 0.9912 and an Expected Calibration Error (ECE) of 0.0566, demonstrating a strong balance between predictive accuracy and confidence reliability. DeepSeek + RAG achieved an F1-score of 0.9989 and an AUC of 0.9970, indicating superior discriminative capability, although with a slightly higher ECE of 0.0634. Relative to the baseline approaches, the proposed model outperforms RoBERTa–base (F1–score = 0.9417, AUC = 0.9565), achieving absolute gains of approximately +5.7% in F1–score and +4.0% in AUC, alongside consistent improvements in Matthews Correlation Coefficient (MCC), Brier Score (BS), Logarithmic Loss (LogLoss), and Precision–Recall AUC (PR–AUC) across all evaluated Reddit categories. These gains, validated via the Wilcoxon signed–rank test ( p < 1 0 − 5 ), confirm the statistical significance and robustness of the approach. Beyond accuracy, LLaMA 3.1 + RAG and DeepSeek + RAG exhibit emergent scaling behavior, with performance increasing non–linearly as training FLOPs grow.