On July 13th, Qifu Technology officially open-sourced ModelEvo, an automated modeling agent designed for business applications.
This tool aims to consolidate dispersed expert knowledge into standardized, reusable modeling capabilities to enhance internal R&D efficiency, while also offering the methodology validated in real-world operations to the industry, contributing to the collaborative development of AI modeling infrastructure.
Users simply need to define their business objectives; ModelEvo can then orchestrate standardized modeling skills via its Agent to assist with requirement clarification, data inspection, evaluation of existing models, sample construction, feature analysis, model training, automated tuning, performance evaluation, and report generation.
The initial version supports two typical tasks: classification prediction and Uplift modeling.
The Challenge of Modeling Beyond Algorithms
Modeling is difficult, and the challenge extends beyond mere algorithm optimization.
Translating business goals into modeling tasks, defining labels, segmenting sample windows, identifying feature leakage, and evaluating incumbent models all directly impact model efficacy and business value.
Historically, these processes were often scattered across individual notebooks and ad-hoc scripts, leading to redundant development and inconsistent standards.
ModelEvo seeks to transform this dispersed expert knowledge into standardized, executable, and traceable modeling capabilities.
Assess First, Build Later for Continuous Asset Reuse
After a user clarifies their requirements, ModelEvo first queries its model knowledge base to retrieve historical models with similar objectives, customer segments, or feature systems.
Based on metrics like AUC, KS, bucket ranking, and business applicability, it determines whether a model can be reused directly, requires further optimization, or needs to be built anew.
This mechanism enhances the reuse rate of historical models and expertise, allowing model assets to generate sustained value.
Translating Expertise into Executable Skills
ModelEvo codifies the methods and quality standards from Qifu Technology's real business operations into rules understandable by its Agent, decomposing the modeling workflow into composable, reusable, and traceable Skills.
The system can check label definitions, observation windows, and performance windows, provide suggestions for feature selection, model choice, and parameter optimization, and automatically generate evaluation results including AUC, KS, and bucket ranking.
It can also perform multiple rounds of self-iterative optimization based on evaluation outcomes, continuously driving improvements in model performance.
The Agent records the data, features, parameters, metrics, and model artifacts for each experiment.
What the user ultimately receives is not just a model, but a complete experiment log, model comparison results, and a reproducible modeling report suitable for professional review.
Lowering Barriers Without Compromising Standards
ModelEvo automates standardized, repetitive tasks, thereby amplifying the value of professional expertise.
Business personnel can participate in problem definition and result interpretation, data analysts can handle data exploration and baseline validation, while algorithm engineers can delegate routine training, data checks, and report generation to the Agent, freeing up more focus for complex scenarios and technological innovation.
Compared to traditional AutoML, which primarily focuses on algorithm selection and parameter search, ModelEvo places greater emphasis on understanding business problems, reusing existing models, and standardizing the complete workflow, conducting multiple rounds of iterative optimization based on evaluation feedback to gradually explore the self-evolution capabilities of features and models.
Wang Yaoxuan, Head of Growth Algorithm at Qifu Technology, stated, "Large language models are rapidly lowering the barriers to code development and the use of algorithmic tools. However, building effective models in real business scenarios still requires deep insight into business problems, professional modeling judgment, and the ability to harness large models. Developing ModelEvo is precisely about consolidating these three types of capabilities and the expert experience behind them into a set of modeling methodologies that are validated in real scenarios, reusable, traceable, and capable of continuous evolution."
From Tool Open-Sourcing to Collaborative Capability Building
ModelEvo v1.0 includes complete examples based on public datasets, allowing users to experience the core workflow locally without deploying a big data cluster; enterprise users can follow the README.md for full-process integration.
In the future, the project will gradually expand capabilities such as the model knowledge base, feature self-evolution, and model self-evolution, continuously refining the intelligent modeling system.
Qifu Technology aims, through open-sourcing, to transform modeling expertise honed in real business into an open, reusable, and sustainably evolving industry capability, promoting the evolution of business modeling from reliance on individual experience towards process standardization, experience capitalization, and capability intelligence.
Comments