arXivPaperNeeds Review

Where A Small Language Model Helps in Invoice Categorisation, Understood Through Embedding Geometry

A single-GPU SBERT beat the reported zero-shot LLM and vendor baseline for invoice coding, suggesting narrow, private classifiers can outperform broader models with modest local data.

arXiv

Open Source Open Markdown Open JSON

Source Summary

A fine-tuned SBERT reached **0.96 accuracy** on invoice classification, above the study’s zero-shot LLM and vendor baseline. For new-client generalization it reached **0.9 F1** with roughly **100 client-specific invoices**, using one GPU.

Practical Implication

For narrow classification with sensitive data, benchmark a small in-house encoder before defaulting to a hosted general model. Inspect embedding clusters and test raw versus human-friendly structured inputs rather than assuming extra formatting helps.

Agent-Ready Context

A fine-tuned SBERT reached **0.96 accuracy** on invoice classification, above the study’s zero-shot LLM and vendor baseline. For new-client generalization it reached **0.9 F1** with roughly **100 client-specific invoices**, using one GPU.

For narrow classification with sensitive data, benchmark a small in-house encoder before defaulting to a hosted general model. Inspect embedding clusters and test raw versus human-friendly structured inputs rather than assuming extra formatting helps.

The results concern one financial corpus and task, so they do not establish a general small-model advantage. Vendor identity strongly shaped the embedding space, and the material does not report deployment cost or performance on other accounting datasets.

Context Map

modeldata#model-selection

Uncertainty

The results concern one financial corpus and task, so they do not establish a general small-model advantage. Vendor identity strongly shaped the embedding space, and the material does not report deployment cost or performance on other accounting datasets.