SuperAPI AI Directory
Partner with us中文AI newsFavorites
All toolsAI financeqwen-dianjin

qwen-dianjin

Qwen DianJin is a collection of open-source projects from the Alibaba Cloud financial services industry team for exploration, research, and deployment of large models in the financial domain. The repository is organized around the financial data flywheel methodology of "evaluation → data synthesis → post-training," and has released achievements including financial evaluation benchmarks, training datasets, reasoning models, process reward models, multimodal document understanding models, and a financial agent skill library. Representative outputs include the CFLUE Chinese financial language understanding evaluation dataset, DianJin-R1-7B/13B and DianJin-R1-Data, the Fin-PRM financial-domain process reward model, DianJin-OCR-R1, and others; it also covers directions such as M³FinMeeting, FinMCP-Bench, DianJin-RED, and DianJin-SKILLS (Baiji Tu). The research is productized on the Alibaba Cloud Qwen DianJin platform. The code and documentation are open under the MIT license, but some components and datasets have separate license terms.

Open website

About this tool

Qwen DianJin is the open-source project homepage of the Alibaba Cloud financial services industry team, focusing on the exploration, research, and deployment of large models in the financial industry, at the address https://github.com/aliyun/qwen-dianjin。

I. Purpose and Positioning This repository serves as an open-source summary of the team's financial AI research, centrally releasing code, models, and data, and synchronizing research progress. Its research is productized in the Qwen DianJin Platform, Alibaba Cloud's AI workspace for financial institutions; specific product information and access methods for the platform must be found on the platform site.

II. Methodology: The DianJin Data Flywheel The team adopts a closed-loop methodology: evaluation → data synthesis → post-training. Evaluation provides quantifiable capability goals through industry-level benchmarks; data synthesis uses controllable, scalable trajectory data to transform the long-standing problem of data scarcity in the financial domain into "data can be produced"; post-training (SFT, RL) transforms data into stable, transferable task capabilities and feeds back into evaluation and data generation.

III. Research Directions

  1. Financial benchmarks and evaluation: a three-tier capability framework covering foundational capabilities (CFLUE), scenario capabilities (M³FinMeeting), and agent capabilities (FinMCP-Bench), with FinGDPVal under joint construction.
  2. Data synthesis: agent- and workflow-based trajectory distillation, MCP environment interaction, and multi-agent self-play, supporting datasets such as DianJin-R1-Data and the CSC corpus.
  3. Post-training: the DianJin-R1 reasoning model and Fin-PRM, the first process reward model in the financial domain.
  4. Intelligent dialogue: customer service conversation (CSC) and emotional support conversation (CARE), with explicit cognition–emotion–strategy reasoning chains.
  5. Agents: the DianJin-SKILLS (Baiji Tu) skill library and DianJin-TIR autonomous planning and tool use.
  6. Multimodal: reasoning and tool-interleaved vision-language models for financial document understanding (DianJin-OCR-R1). The key directions for 2026 are agentic reinforcement learning (Agentic RL) and meta-harness evolution.

IV. Open-Source Outputs and Timeline (based on publicly available repository information)

  • 2024.05.16: The CFLUE-related paper was accepted by ACL 2024.
  • 2025.01.06: The CFLUE dataset was fully open-sourced and made available for download.
  • 2025.04.23: The DianJin-R1 series was open-sourced, including the DianJin-R1-Data dataset and the two models DianJin-R1-7B and DianJin-R1-13B; see the technical report "DianJin-R1: Evaluating and Enhancing Financial Reasoning in Large Language Models》.
  • 2025.05.22: M³FinMeeting (a multilingual, multi-industry, multi-task financial meeting understanding evaluation dataset) was accepted by ACL 2025.
  • From 2025.08.08: research results related to customer service conversations (Evaluating, Synthesizing, and Enhancing for Customer Support Conversation) were released and open-sourced.
  • 2025.08.18: DianJin-OCR-R1 was released and open-sourced.
  • 2025.08.25: Fin-PRM was released and open-sourced.
  • 2025.10.11: FinMCP-Bench was jointly released by Yingmi Fund and others, as the first benchmark dataset and evaluation framework for real financial tool use by MCP-based LLM agents; CARE was released during the same period.
  • 2025.11.15: the paper related to customer service conversations was officially accepted by AAAI 2026.
  • 2026.02.27: the FinMCP-Bench and CARE papers were accepted by ICASSP 2026.
  • 2026.04.07: the CFMME-related paper was accepted by the ACL 2026 main conference.
  • 2026.05.10: the Fin-PRM paper was accepted by IJCAI 2026.
  • 2026.05.20: DianJin-SKILLS (Baiji Tu) was open-sourced, as an AI Agent skill library covering banking, insurance, and securities/asset management, including 10 professional roles and 130+ standardized skills, and can be directly integrated into Agent frameworks.
  • 2026.08.17: DianJin-RED was released, as an action-anchored red-teaming benchmark for complete agent systems, containing 1,661 executable test cases, 15 intervention strategies (user input, agent-platform state, external tools/data), and an isolated service environment.
  • 2026.08.25: the FinGuard-related paper was accepted by EMNLP 2026 Findings.

V. Collaboration and Co-construction The team, together with industry partners, is building FinGDPVal to evaluate whether an AI workspace can complete real financial position tasks end to end; its gold-standard evaluation set and toolchain will be open-sourced in this repository. Researchor product inquiries can be made by contacting CFLUE@alibabacloud.com, or by scanning the QR code to join the DingTalk group. The project welcomes contributions in accordance with CONTRIBUTING.md and the code of conduct.

VI. License and Usage Restrictions Qwen DianJin is developed by Alibaba Cloud. The copyright of the original code and documentation in the repository belongs to Alibaba Cloud and its affiliates (2026), and is licensed under the MIT License, unless otherwise stated in a file or subdirectory. The repository contains third-party source code and separately licensed data, and existing file-level and component-level copyright and license notices must be retained; among them, part of the content of DianJin-PRM and DianJin-OCR-R1 originates from Apache-2.0 projects, the DianJin-RED benchmark data is subject to CC BY-NC 4.0 and related upstream terms, and the MIT License in the root directory does not cover these terms; see NOTICE for details. The team states that it assumes no legal liability for the use of the DianJin open-source models and data; users must evaluate and bear potential risks themselves, verify model outputs, and make informed decisions; the released data and models are intended for academic research and industry applications, to advance AI development in data analysis, financial innovation, and related fields.

VII. Parts Not Provided in the Materials The repository text does not provide information such as the platform's specific feature list, charging methods and prices, usage step-by-step tutorials, or model performance comparison figures, and these are not provided for now.

Related tools

Browse category