Researchers led by Alina Khaybullina released FinVector-Market-4B, a version of Qwen3.5‑4B that was adapted with rank‑16 LoRA on 22,000 financial examples. The study compares the base model and the adapted one on a 600‑example benchmark using JSON‑schema prompts. Providing only the schema lifts JSON validity from 0% to 91.3% for the base model. With matched explicit prompting, the adapted model raises FinQA exact‑match accuracy from 14.7% to 40.0% and improves calculator‑expression correctness from 48.0% to 82.7%.
Why it matters
The gains suggest small, targeted fine‑tuning can make general‑purpose models much more useful for finance professionals.
