How to Fine-Tune an SLM for Emotion Recognition

This tutorial details how to fine-tune a Mistral Small 3.1 language model for multi-label emotion recognition using an imbalanced dataset. To address the data imbalance in the GoEmotions dataset, the process combines undersampling the majority class, synthetically creating new samples for minority classes with the ISMOTE algorithm, and using a weighted focal loss function. The article provides Python code snippets for implementation, leveraging the Unsloth framework for efficient training with LoRA. The resulting model is designed to classify 15 specific emotions from text with improved performance on underrepresented categories. The final fine-tuned model has been made available on Hugging Face.