Parameter-Efficient Fine-Tuning of Compact Language Models for Professional LinkedIn Post Generation

Abstract

The rise of generative AI has transformed social media content creation, but professional networking platforms such as LinkedIn demand domain-specific, contextually accurate, and stylistically appropriate outputs. Existing large language model (LLM) solutions are often computationally expensive and lack specialization for professional contexts. This paper presents a lightweight, parameter-efficient approach for LinkedIn post generation using Low-Rank Adaptation (LoRA) applied to TinyLlama-1.1B-Chat. A curated dataset of 500 job-related instruction–output pairs was extracted, preprocessed, and fine-tuned in a Google Colab environment using Hugging Face Transformers and PEFT libraries. Evaluation combined lexical similarity via difflib.SequenceMatcher with precision, recall, and F1-scores against human-curated references. The fine-tuned model achieved an average similarity of 0.78 and correctness in 85% of cases, with precision, recall, and F1 all at approximately 75%. LoRA reduced trainable parameters to 8.8M (0.8% of the full model), cutting memory usage and training time by about 70% while maintaining stable training curves and robust generalization to unseen prompts. Generated posts were qualitatively assessed as professional, coherent, and contextually relevant. These results demonstrate the feasibility of creating resource-efficient, domain-specific generators for professional communication. The proposed pipeline offers a replicable and scalable framework for automating LinkedIn post creation, bridging the gap between general-purpose LLMs and specialized professional content generation.

Ähnliche Arbeiten