Question
Can a small locally running model classify transactions reliably enough for production use?
LAB 001 / EXPERIMENT
Local LLM Transaction Classification
Testing whether small local models can reliably classify financial transactions.
01 / LAB NOTES
Hypothesis
A compact model should handle obvious merchant categories well, while ambiguous transfers need deterministic rules or human review.
Experiment
Compare zero-shot prompts, few-shot prompts, and a hybrid rules-plus-model pipeline on a labeled transaction set.
Result
The hybrid approach produced the clearest confidence boundary and the most actionable error set.
What I changed
Moved merchant normalization before inference and added an explicit unknown category instead of forcing a guess.
TOOLS & TECHNOLOGIES
01Python02Ollama03Qwen