AI BriefingOptimizing Small Model Performance Using LLMs
Key point
This introduces CFM's case of using LLM-based data labeling to boost small model performance while cutting operational costs by up to 80x.
Details
Asset management firm CFM built a workflow combining LLM-based data labeling with small model fine-tuning to accurately extract entities such as company names from financial news.
The conventional approach of directly using large-scale LLMs to perform NER (Named Entity Recognition) offers high accuracy but has limitations in terms of cost and scalability. To address this, CFM went through the following steps.
- LLM-based labeling: Deploy Llama 3.1-70b-Instruct on Hugging Face Inference Endpoints to generate high-quality synthetic labeled data.
- Data cleaning: Use Argilla to review the generated predictions and improve data quality.
- Small model fine-tuning: Fine-tune lightweight models such as GLiNER and SpanMarker based on the cleaned data.
As a result, GLiNER's F1 score improved from 87.0% to 93.4%, and SpanMarker saw a substantial improvement from 47.0% to 90.1%. In particular, in terms of operational cost, it demonstrated efficiency by achieving similar performance levels at up to 80x lower cost compared to large-scale LLMs.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Qwen3.5-based Small Decision Model 'Kev' Released with TypeSafe API Compatibility and Performance Metrics