LLM vs Fine-Tuning

• Python 3.11 / 3.12
• PyTorch
• Hugging Face Transformers
• DistilBERT
• scikit-learn (TF-IDF, LinearSVC)
• Anthropic Claude API
• Structured Output / Tool Use
• QLoRA / PEFT
• Qwen2.5-1.5B
• FastAPI
• Gradio
• Hugging Face Spaces
• Docker
• pytest (92 tests, CI)
• pydantic-settings

Empirical study that reframes emotion classification as a production decision rather than a leaderboard contest. I benchmarked four approaches on the same 417k-sample corpus: TF-IDF and Linear SVM, fine-tuned DistilBERT, prompted Claude Haiku 4.5 with constrained structured output (zero- and few-shot), and QLoRA on Qwen2.5-1.5B. Each was measured on macro-F1, p95 latency and dollar cost per 1,000 predictions. The headline result: the 67M-parameter fine-tuned encoder reaches 0.9128 macro-F1 and beats the frontier LLM by roughly 40 points at effectively zero per-prediction cost, while even the classical SVM beats the LLM by 32 points. Shipped with a live three-way classifier on Hugging Face Spaces, a FastAPI inference service, a Docker image, a 92-test reproducibility harness and a written paper.