Open Access||Research ArticleFinistral AI: An Efficient Financial-Sentiment LoRA Adapter and a Train/Test Contamination Case Study
1 Indiana University Bloomington, 107 S. Indiana Avenue, Bloomington, IN 47405-7000, USA
2 Tata Consultancy Services (TCS), TCS House, Raveline Street, Fort, Mumbai, Maharashtra, 400001, India
3 Arizona State University, 1151 S Forest Ave, Tempe, AZ, United States
4 University of Mumbai, Mumbai, Maharashtra, 400032, India
Abstract
We present Finistral-7B-LoRA, a parameter-efficient financial-sentiment classifier produced by applying Low-Rank Adaptation (LoRA) to the 7-billion-parameter Mistral-7B-v0.1 language model. The adapter is trained on the open-source FinGPT/fingpt-sentiment-train corpus of approximately 77,000 labelled sentences, updating ≈ 0.58% of the model parameters (≈ 41.9 M trainable, matching the released adapter's configuration), on two NVIDIA A100 (40 GB) GPUs of Indiana University's Big Red 200 supercomputer. On the full Financial Phrase Bank sentences all agree split the adapter shows an apparent 99.56% accuracy, but our own data-overlap audit reveals that 75.2% of those evaluation sentences appear verbatim, with identical gold labels, inside the training corpus, so that figure reflects memorization rather than generalization and is retained only as a contamination diagnostic. The paper's primary evidence is instead a corrected, exact-match-decontaminated evaluation: on the 560-sentence decontaminated Financial Phrase Bank remainder Finistral attains 98.9% accuracy, on FiQA-SA 87.7% accuracy / 0.883 weighted F1, and on Twitter Financial News Sentiment 78.9% accuracy / 0.796 weighted F1. These are statistically indistinguishable from the strongest FinGPT-class adapter on each dataset and significantly ahead of FinBERT off-distribution, with all comparisons carrying per-example significance tests. Our main contributions are (i) an efficient, fully reproducible single-node LoRA recipe for financial sentiment whose adapter runs inference on a single GPU, and (ii) a cautionary evaluation/contamination case study quantifying how corpus overlap and evaluation-harness defects can simultaneously inflate a model and deflate its baselines in the financial-NLP literature. The released adapter (Ayansk11/Finistral-7B_lora on Hugging Face), decontamination and evaluation scripts, and frozen evaluation sets make every reported evaluation number independently reproducible.
Keywords
Graphical Abstract
Novelty Statement
Introduces Finistral-7B-LoRA as an efficient, reproducible financial-sentiment PEFT approach, while establishing contamination auditing and evaluation-harness validation as essential requirements for credible LLM benchmarking.