Financial Large Language Model: Characteristics, Construction, Architecture, and Practical Applications

The presentation introduces Qifu Technology’s AI‑driven credit service platform and outlines four main topics: the features of financial large models, how to build them, the technical framework and process, and practical deployment scenarios.

Financial Large Model Characteristics – The finance industry has evolved from paper to digital, then to big‑data and deep‑learning eras, and now to large‑model‑assisted decision making. Large models act like a high‑speed vehicle where business understanding is the cockpit, compliance is the steering wheel, private domain data is fuel, logical reasoning is the engine, and compute resources are the tires.

How to Build Financial Large Models – Two major aspects are highlighted: “soft” strengths (deep business insight and strict compliance) and “hard” strengths (rich private data, diverse technical reserves, and sufficient compute). Data processing includes filtering low‑quality data, de‑duplication, quality control, privacy removal, tokenization, and ordering.

The technical stack is described in five layers:

Data Layer : combines private financial data (graphs, dialogues, behavior logs, reports) with relevant public data.

Data Engineering Layer : handles data cleaning, ratio balancing, and prompt engineering for tasks such as sentiment analysis, reading comprehension, and domain‑specific Q&A.

Model Layer : a L1 financial model fine‑tuned for top‑level applications like risk control and telemarketing.

Evaluation Layer : assesses data quality, data‑engineering effectiveness, model performance on both general and finance‑specific benchmarks, and business impact.

Application Layer : implements the model in concrete scenarios.

Technical Process – Consists of three stages: constructing a large‑scale finance language model, injecting general and finance‑specific capabilities, and fine‑tuning for specific tasks (e.g., risk management, intent recognition, advertising).

Evaluation Scheme – Combines subjective and objective questions of varying difficulty (high‑school, undergraduate, professional) and mixes manual, automated, and large‑model‑assisted assessments, shifting the balance as the model matures.

Practical Deployments :

Quality inspection: using the model to analyze de‑identified call recordings, achieving a 15% detection improvement and 100% coverage.

AIGC material generation: automated creation and transformation of advertising assets, boosting reach by 21.4%.

Sales‑lead mining: extracting high‑value cues from call data, raising lead‑recognition accuracy to 98% and improving conversion by 4.6%.

Intelligent Q&A with knowledge graphs: combining chain‑of‑thought prompting, graph query generation, and model‑based answer synthesis for precise, natural responses.

The Q&A section addresses model training for material generation, the role and reliability of intent recognition in finance, and methods for extracting key information for lead mining.

Overall, the financial large model serves as a specialized expert assistant that augments, rather than replaces, human decision‑makers in the highly regulated finance domain.