Why use a small language model for sentiment analysis?

Small language models can be fine-tuned on a company's own examples to understand its terminology and customer language, often at lower latency and cost than large general-purpose models. Whether that holds for a given workload is confirmed by comparison against baseline approaches.

How does the model handle mixed or unclear sentiment?

The labeling rules include a mixed category, and the model is evaluated specifically on mixed-opinion examples. Uncertain predictions are routed to human review rather than forced into a confident answer.

What is continuous learning in this context?

It is a controlled retraining loop. Reviewers correct the model's mistakes, verified corrections are added to the training data, and new model versions are trained and tested before release.

How do you stop a model update from getting worse?

Every candidate is tested against the current model on recent unseen data and a protected regression test set. It is only released if it meets the agreed criteria, and rollback is built into the workflow.

Why keep model predictions separate from training labels?

Training a model on its own unverified predictions can reinforce its mistakes. Only human-reviewed labels enter the training dataset.

What does it cost to keep the model up to date?

The recommendation accounts for the full ongoing cost: labeling, human review, retraining, and inference, not inference alone.

Small Language Model Fine-Tuning for Sentiment Analysis with Continuous Learning - Case Study

Executive Summary

Customer feedback is rarely just positive or negative: a customer may praise the support agent while describing an unresolved problem, and a polite message can still express frustration. GenAI Protos fine-tunes a small language model on the client's own examples to read that feedback more accurately, and builds a feedback loop where reviewed corrections train future versions. Every update is tested against the current model and only released if it improves performance without losing accuracy on previously understood cases.

01Client Requirement

The client needs to classify customer feedback across reviews, surveys, or support conversations within one business domain. The challenge is understanding company-specific language, mixed opinions, and changing customer concerns, along with a practical way to correct recurring mistakes as new products and terminology appear.

Apply consistent sentiment labels using the client's definitions
Recognize negative feedback even when it is expressed indirectly
Handle mixed or unclear sentiment without forcing a confident answer
Process the expected volume within agreed latency and cost limits
Incorporate reviewed corrections into future training
Check every model update before it replaces the current version

02Our Approach

We begin with the labeling rules. The client's reviewers agree on what positive, negative, neutral, and mixed sentiment mean for the task, using examples where interpretations differ. We then prepare de-identified training, validation, and test datasets, with the test data including difficult language and more recent feedback to check whether the model handles examples beyond its training history.

A conventional sentiment classifier, a prompted language model, and the selected small model provide the initial comparison. We fine-tune the small model on the agreed examples and measure:

Accuracy by sentiment class
Missed negative feedback
Errors involving mixed opinions or domain-specific language
Response time, throughput, and processing cost

03Outcome

A fine-tuned model or adapter and a working inference setup
A comparison with the baseline approaches, including remaining failure cases
Prepared datasets, labeling guidance, and evaluation assets
A feedback and retraining workflow with release checks and rollback
Findings on accuracy, review effort, latency, and operating cost

Small Language Model Fine-Tuning for Sentiment Analysis with Continuous Learning