Paper: arXiv 2511.08621
Authors: Gaëtan Caillaut, Raheel Qader, Jingshu Liu, Mariam Nakhlé, Arezki Sadoune, Massinissa Ahmim, Jean-Gabriel Barthelemy
Abstract
The financial industry’s growing demand for advanced natural language processing (NLP) capabilities has highlighted the limitations of generalist large language models (LLMs) in handling domain-specific financial tasks. To address this gap, we introduce the LLM Pro Finance Suite, a collection of five instruction-tuned LLMs (ranging from 8B to 70B parameters) specifically designed for financial applications. Our approach focuses on enhancing generalist instruction-tuned models, leveraging their existing strengths in instruction following, reasoning, and toxicity control, while fine-tuning them on a curated, high-quality financial corpus comprising over 50% finance-related data in English, French, and German. We evaluate the LLM Pro Finance Suite on a comprehensive financial benchmark suite, demonstrating consistent improvement over state-of-the-art baselines in finance-oriented tasks and financial translation. Notably, our models maintain the strong general-domain capabilities of their base models, ensuring reliable performance across non-specialized tasks. This dual proficiency, enhanced financial expertise without compromise on general abilities, makes the LLM Pro Finance Suite an ideal drop-in replacement for existing LLMs in financial workflows, offering improved domain-specific performance while preserving overall versatility. We publicly release two 8B-parameters models to foster future research and development in financial NLP applications: https://huggingface.co/collections/DragonLLM/llm-open-finance.
Complexity vs Empirical Score
- Math Complexity: 1.5/10
- Empirical Rigor: 8.0/10
- Quadrant: Street Traders — empirical and applied, lower math complexity
Why this score: The paper’s primary contribution is empirical, involving data curation, model fine-tuning on a large, structured dataset, and evaluation on a comprehensive financial benchmark suite, with links to released models. The mathematics is largely conceptual, focusing on LLM architectures and training methodologies rather than dense analytical derivations.
Research Flowchart
flowchart TD A["Research Goal<br>Develop specialized LLMs for finance tasks<br>without losing general capabilities"] --> B["Methodology<br>Instruction-tune 5 LLMs (8B-70B) on curated corpus"] B --> C["Data & Inputs<br>High-quality financial corpus<br>English, French, German"] C --> D["Computational Process<br>Fine-tuning & Instruction Tuning"] D --> E["Evaluation<br>Benchmark suite vs. SOTA baselines"] E --> F["Key Outcomes<br>1. Consistent improvement in financial tasks<br>2. Retained general-domain performance<br>3. Released 8B models for research"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper introduces the LLM Pro Finance Suite, a collection of five instruction‑tuned large language models ranging from 8B to 70B parameters for financial applications.
Source sentence
To address this gap, we introduce the LLM Pro Finance Suite, a collection of five instruction‑tuned LLMs (ranging from 8B to 70B parameters) specifically designed for financial applications.
Assumes
No explicit assumptions found in the text.
Data
- curated financial corpus text-news unknown
Source sentence
fine‑tuning them on a curated, high‑quality financial corpus comprising over 50 % finance‑related data in English, French, and German.
- SFT dataset synthetic unknown
Source sentence
The SFT dataset combines open‑source resources from Hugging Face Hub with synthetic data generated from text documents.
- in‑house financial translation dataset text-news unknown
Source sentence
We evaluated the translation capabilities of LLM Pro Finance models using an in‑house translation dataset covering 8 languages and 22 directions.
Costs
Not applicable
Finds 5
- LLM Pro Finance models achieve consistent improvement over state‑of‑the‑art baselines on finance‑oriented tasks and financial translation, while maintaining general‑domain performance. performance
Source sentence
We evaluate the LLM Pro Finance Suite on a comprehensive financial benchmark suite, demonstrating consistent improvement over state‑of‑the‑art baselines in finance‑oriented tasks and financial translation.
- On non‑financial benchmarks the Pro Finance models are generally on par with their baseline models, indicating no degradation of general‑domain knowledge. performance
Source sentence
The LLM Pro Finance models are generally on par with their baseline models, indicating that general domain knowledge and natural language understanding capabilities are not degraded through the adaptation to the financial domain.
- Pro Finance models outperform baselines in understanding financial acronyms, with accuracy gains ranging from ~10 % to >60 % across languages. performance
Source sentence
The Pro Finance models consistently out‑perform their baseline
- Financial translation performance improves significantly, with an average BLEU improvement of around 15 % across all directions for the three largest models. performance
Source sentence
The performance of all models improved significantly, with an average improvement around 15 % overall. We also show the absolute BLEU
- Training on the same dataset increased toxicity for two models while decreasing it by more than 50 % for two others, showing that data filtering alone does not guarantee safety. negative
Source sentence
we were surprised to observe that training on the same dataset resulted in a substantial increase in toxicity for two models, while leading to more than a 50 % decrease for two others.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 31 related papers). Coverage is partial, so this is neither replication nor refutation.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (40)
- Qwen3 technical report (2025) — link
- The llama 3 herd of models (2024) — link
- Bloomberggpt: A large language model for finance (2023) — link
- Pixiu: A large language model, instruction data and evaluation benchmark for finance (2023) — link
- FinGPT: Open-Source Financial Large Language Models — → 2306.06031
- Finbert: A pretrained language model for financial communications (2020) — link
- Finbert: Financial sentiment analysis with pre-trained language models (2019) — link
- Finbert: A pre-trained financial language representation model for financial text mining (2021)
and 32 more
- FinBERT: A large language model for extracting information from financial text (2023)
- Finben: A holistic financial benchmark for large language models (2024)
- Demystifying domain-adaptive post-training for financial llms (2025) — link
- FineWeb-Edu: the Finest Collection of Educational Content (2024) — link
- The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale (2024) — link
- FineWeb2: One Pipeline to Scale Them All – Adapting Pre-Training Data Processing to Every Language (2025) — link
- Effect of scale on catastrophic forgetting in neural networks (2021)
- Investigating the catastrophic forgetting in multimodal large language models (2023) — link
- An empirical study of catastrophic forgetting in large language models during continual fine-tuning (2025)
- Catastrophic forgetting in deep learning: A comprehensive taxonomy (2023) — link
- DeBERTaV3: Improving DeBERTa using ELECTRA-Style Pre-Training with Gradient-Disentangled Embedding Sharing (2021) — link
- DEBERTA: DECODING-ENHANCED BERT WITH DISENTANGLED ATTENTION (2021) — link
- Rubric is all you need: Enhancing llm-based code evaluation with question-specific rubrics (2025) — link
- News from OPUS - A Collection of Multilingual Parallel Corpora with Tools and Interfaces (2009)
- Qwen3Guard Technical Report (2025)
- Detoxify (2020)
- Bert: Pre-training of deep bidirectional transformers for language understanding (2019)
- Llama: Open and efficient foundation language models (2023) — link
- Fineval: A chinese financial domain knowledge evaluation benchmark for large language models (2023) — link
- FinQA: A Dataset of Numerical Reasoning over Financial Data (2021)
- ConvFinQA: Exploring the Chain of Numerical Reasoning in Conversational Finance Question Answering (2022)
- FinanceBench: A New Benchmark for Financial Question Answering (2023) — link
- HellaSwag: Can a Machine Really Finish Your Sentence? (2019)
- Mmlu-pro: A more robust and challenging multi-task language understanding benchmark (2024)
- Training Verifiers to Solve Math Word Problems (2021) — link
- Ragas: Supercharge Your LLM Application Evaluations (2024)
- Zero: Memory optimizations toward training trillion parameter models (2020)
- TRL: Transformer Reinforcement Learning (2020)
- Bleu: a method for automatic evaluation of machine translation (2002)
- Scaling Laws of Decoder-Only Models on the Multilingual Machine Translation Task (2024) — link
- Large language model adaptation for financial sentiment analysis (2024) — link
- The Language Model Evaluation Harness (2024) — link
1 of 40 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.