deepanalyze

DeepAnalyze, developed by teams from Renmin University of China and Tsinghua University, is positioned as an agentic large language model (LLM) for autonomous data science. Its goal is to complete end-to-end data science tasks without human intervention, including data preparation, analysis, modeling, visualization, and report generation, while supporting open-ended deep research across multiple data sources. Key Capabilities - Covers the full data science pipeline: Automatically executes tasks such as data preparation, analysis, modeling, visualization, and report generation. - Open-ended data research: Handles structured data (databases, CSV, Excel), semi-structured data (JSON, XML, YAML), and unstructured data (TXT, Markdown), ultimately producing analytical reports. - Fully open-sourced: The model (DeepAnalyze-8B), code, training data (DataScience-Instruct-500K), and demos are all open-sourced, facilitating local deployment or further development. Usage Methods - The model can be deployed via vLLM, offering an OpenAI-style API, along with a memory configuration recommendation table covering scenarios from 16GB to 80GB VRAM, supporting quantized models and FP8 KV Cache. - Provides multiple interfaces: WebUI, WebUI v2 (supporting HeyWhale API and Docker-based sandboxed code execution), JupyterUI, and a command-line interface (supporting Chinese and English). - Offers API key application channels and usage documentation, and also supports building custom OpenAI-style services. Training and Evaluation - Supports fine-tuning based on DeepSeek-R1-0528-Qwen3-8B or DeepAnalyze-8B, providing scripts for single-capability fine-tuning, multi-capability agent cold-start training, and reinforcement learning training. - Unifies evaluation processes for most existing data science benchmarks in the playground, facilitating assessment of DeepAnalyze or custom agents. Ecosystem and Related Projects - Related projects include SkillAdam (an agent skill auto-optimization plugin), DeepPrep (a data preparation agent), CoDA-Bench (a code agent benchmark for data-intensive analysis tasks), and DA-Studio (the system behind WebUI v2). - The training framework is based on ms-swift and SkyRL, with training data sourced from Reasoning-Table, Spider, BIRD, DABStep, and others. Applicable Scenarios Suitable for users and teams needing automated data analysis, batch generation of analytical reports, or local private deployment of data analysis assistants.