QianfanHuijin Technical Report: A Novel Multi-Stage Training Paradigm for Finance Industrial LLMs

Domain-specific enhancement of Large Language Models (LLMs) within the financial context has long been a focal point of industrial application. While previous models such as BloombergGPT and Baichuan-Finance primarily focused on knowledge enhancement, the deepening complexity of financial services has driven a growing demand for models that possess not only domain knowledge but also robust financial reasoning and agentic capabilities. In this paper, we present QianfanHuijin, a financial domain LLM, and propose a generalizable multi-stage training paradigm for industrial model enhancement. Our approach begins with Continual Pre-training (CPT) on financial corpora to consolidate the knowledge base. This is followed by a fine-grained Post-training pipeline designed with increasing specificity: starting with Financial SFT, progressing to Finance Reasoning RL and Finance Agentic RL, and culminating in General RL aligned with real-world business scenarios. Empirical results demonstrate that QianfanHuijin achieves superior performance across various authoritative financial benchmarks. Furthermore, ablation studies confirm that the targeted Reasoning RL and Agentic RL stages yield significant gains in their respective capabilities. These findings validate our motivation and suggest that this fine-grained, progressive post-training methodology is poised to become a mainstream paradigm for various industrial-enhanced LLMs.

Learning from Training Dynamics: Identifying Mislabeled Data Beyond Manually Designed Features

While mislabeled or ambiguously-labeled samples in the training set could negatively affect the performance of deep models, diagnosing the dataset and identifying mislabeled samples helps to improve the generalization power. Training dynamics, i.e., the traces left by iterations of optimization algorithms, have recently been proved to be effective to localize mislabeled samples with hand-crafted features. In this paper, beyond manually designed features, we introduce a novel learning-based solution, leveraging a noise detector, instanced by an LSTM network, which learns to predict whether a sample was mislabeled using the raw training dynamics as input. Specifically, the proposed method trains the noise detector in a supervised manner using the dataset with synthesized label noises and can adapt to various datasets (either naturally or synthesized label-noised) without retraining. We conduct extensive experiments to evaluate the proposed method. We train the noise detector based on the synthesized label-noised CIFAR dataset and test such noise detector on Tiny ImageNet, CUB-200, Caltech-256, WebVision and Clothing1M. Results show that the proposed method precisely detects mislabeled samples on various datasets without further adaptation, and outperforms state-of-the-art methods. Besides, more experiments demonstrate that the mislabel identification can guide a label correction, namely data debugging, providing orthogonal improvements of algorithm-centric state-of-the-art techniques from the data aspect.