How to Fine-Tune an SLM for Emotion Recognition

2026-06-05 · Source: Towards Data Science · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Natural Language Processing · Depth: Intermediate, medium

Summary

A new fine-tuned small language model, MistralSmall-3.1.GoEmotions, has been released on Hugging Face under the Apache 2.0 license, designed for detailed emotion recognition. This model extends traditional sentiment analysis by assigning probabilities across 27 distinct emotion categories, such as "anger" and "disappointment", from text. Developed to address the significant class imbalance in the 58k-entry GoEmotions dataset, particularly the prevalence of "neutral" labels, the fine-tuning process employed a three-pronged strategy. This involved undersampling majority classes, synthetically expanding minority emotion categories using the 2025 ISMOTE algorithm, and implementing a weighted focal loss function. The fine-tuning of Mistral Small-3.1-24B-Instruct-2503, utilizing the Unsloth framework and LoRA, took 9 hours and 30 minutes on an NVIDIA RTX 6000 GPU. The resulting model achieved F1 scores exceeding 0.7 for most target emotions.

Key takeaway

For AI Engineers developing fine-grained emotion recognition models, you must prioritize robust data preprocessing to counter class imbalance. Implement a strategy combining undersampling, synthetic data generation like ISMOTE, and weighted focal loss during fine-tuning. This method, proven with Mistral Small, improves performance on minority emotion categories. It enables more accurate analysis for customer support or brand monitoring.

Key insights

Fine-tuning SLMs for fine-grained emotion recognition on imbalanced datasets requires a multi-technique approach to data balancing and loss weighting.

Principles

• Class imbalance biases models towards majority classes.
• Emotion recognition extends sentiment analysis.
• Open-weight SLMs offer transparency and cost benefits.

Method

Address class imbalance by undersampling majority classes, synthetically expanding minority classes with ISMOTE, and applying a weighted focal loss function during LoRA fine-tuning with Unsloth.

In practice

• Monitor brand sentiment in social media.
• Categorize customer support emails by emotion.
• Analyze emotional content in online discussions.

Topics

• Emotion Recognition
• Small Language Models
• Fine-tuning
• Class Imbalance
• ISMOTE Algorithm
• Focal Loss

Code references

• google-research/google-research
• PetrKorab/MistralSmall-3.1.GoEmotions

Best for: Machine Learning Engineer, NLP Engineer, AI Engineer

Related on AIssential

• I Fine-Tuned DistilBERT Twice. It Humbled Me Both Times. — Data characteristics fundamentally determine fine-tuning performance and metric interpretation, not just the model.
• 30% of Google's Emotions Dataset is Mislabeled — Poor data labeling methodology, especially lacking context and cultural fluency, severely compromises dataset quality and model performance.
• Model-Agnostic Meta Learning for Class Imbalance Adaptation — HAMR dynamically addresses class imbalance and data difficulty in NLP via adaptive weighting and neighborhood-aware resampling.
• How I Fine-Tuned DistilBERT to Classify Complaints — And What I Learned Along the Way — Deeply understanding transformer fine-tuning components and their interconnections is crucial for effective model development.
• On Emotion-Sensitive Decision Making of Small Language Model Agents — Emotion-sensitive decision-making in SLMs is unstable and often misaligned with human expectations, despite systematic influence.
• Portuguese Sentiment Analysis with Open-Source LLMs: Models, Prompts, and Efficient Deployment — Open-source LLMs offer competitive Portuguese sentiment analysis with efficient deployment, especially using CoT+FS prompting.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Towards Data Science.