AI vs. Human Text Detection

A comparative study in Natural Language Processing focused on identifying AI-generated text. This project involved fine-tuning multiple architectures using Parameter-Efficient Fine-Tuning (PEFT) and analyzing model decision-making through interpretability frameworks.

Technical Stack

• Primary Model: RoBERTa-base (Fine-tuned)
• Benchmark Models: Llama 3.2 (1B), DistilBERT-base-uncased
• Fine-Tuning: LoRA (PEFT), Hugging Face Transformers
• Interpretability: LIME (Local Interpretable Model-agnostic Explanations)
• Analysis: McNemar Statistical Testing, Python, Flask

Core Capabilities

• Targeted Fine-Tuning: Executed LoRA (Low-Rank Adaptation) on RoBERTa-base, optimizing only 0.71% of total parameters (887k) to achieve near-state-of-the-art performance with minimal hardware requirements.
• Comparative Benchmarking: Evaluated model performance across 480k records, identifying RoBERTa as the optimal model for production due to a 4x faster training speed compared to Llama 3.2 while maintaining a high 96.7% F1-score.
• Model Interpretability: Integrated LIME explanations to visualize linguistic feature importance, validating that the fine-tuned model correctly identified structural patterns in AI text rather than over-fitting on specific keywords.
• Statistical Validation: Applied McNemar tests to establish that RoBERTa significantly outperformed the DistilBERT baseline, proving the efficacy of the fine-tuning approach for complex text classification.