ruc-datalab/DeepAnalyze

Authors: Shaolei Zhang, Ju Fan*, Meihao Fan, Guoliang Li, Xiaoyong Du

Renmin University of China, Tsinghua University

DeepAnalyze is the first agentic LLM for autonomous data science. It can autonomously complete a wide range of data-centric tasks without human intervention, supporting:

• 🛠 Entire data science pipeline: Automatically perform any data science tasks such as data preparation, analysis, modeling, visualization, and report generation.
• 🔍 Open-ended data research: Conduct deep research on diverse data sources, including structured data (Databases, CSV, Excel), semi-structured data (JSON, XML, YAML), and unstructured data (TXT, Markdown), and finally produce analyst-grade research reports.
• 📊 Fully open-source: The model, code, training data, and demo of DeepAnalyze are all open-sourced, allowing you to deploy or extend your own data analysis assistant.

• [2026.09]: We release EvoOntology, a self-evolving ontology layer for data agents that grounds semantic context in real workloads and continuously adapts through verified execution trajectories.

  Details

  🔗 Demo:

  v3.mp4

• [2026.09]: We release SkillAdam 🚀, an easy-to-use plugin for automatically improving the skills of your AI agents. SkillAdam integrates directly with Codex, Claude Code, Cursor Agent, and GitHub Copilot, with simple installation and no need for a local GPU, Docker, or a separate model API key.

  More about SkillAdam

  SkillAdam makes skill optimization simple and automatic. Just provide a SKILL.md and describe what you want to improve. SkillAdam automatically generates relevant tasks, evaluates the current skill, proposes improvements, validates the revised version, and updates the skill when the changes pass evaluation. With a one-step installer, it works directly inside the coding agent you already use.

  🔗 Demo:

  skilladam-demo.mp4

• [2026.07]: We look forward to releasing DeepPrep, a data-preparation companion to DeepAnalyze that turns raw tables into analysis-ready data.

  More about DeepPrep

  DeepPrep is an LLM-powered agentic system for autonomous data preparation. It constructs data-preparation pipelines through execution-grounded interaction with intermediate table states and runtime feedback, helping clean, transform, and standardize raw data before downstream analysis.

  🔗 Demo:

  deepprep.mp4

• [2026.06.15]: We release CoDA-Bench, a benchmark for evaluating whether code agents can handle data-intensive analytical tasks, closely aligned with DeepAnalyze's target scenarios.

  More about CoDA-Bench

  CoDA-Bench evaluates agents in a Linux sandbox with hundreds of data files. Given a natural-language question, an agent must discover relevant data, write executable code, and produce the final answer. It provides a benchmark setting for the same type of data discovery and code-execution challenges targeted by DeepAnalyze.

  🔗 Demo:

  demo_video_compressed.mp4

• [2026.05.31]: DA-Studio, the system behind DeepAnalyze WebUI v2 (demo/chat_v2), has been accepted to the VLDB 2026 Demonstration Track.

• [2026.03.16]: Update DeepAnalyze WebUI v2, featuring a smoother UI, support for the HeyWhale API, and support for Docker-based sandboxed code execution. More details in Readme .

• [2026.01.31]: 🎉🎉🎉DeepAnalyze served as the official agent supporting the 2026 (19th) China Collegiate Computer Design Contest — Big Data Track.

• [2025.12.28] ANNOUNCEMENT: DeepAnalyze API Keys Are Now Available 🎉🎉🎉 You can now apply for your API key via this Google Form or this Feishu Form. For full details and usage instructions, please refer to the Guide or the Feishu Wiki.

• [2025.11.13]: DeepAnalyze now supports OpenAI-style API endpointsis and is accessible through the Command Line Terminal UI. Thanks to the contributor @LIUyizheSDU

• [2025.11.08]: DeepAnalyze is now accessible through the JupyterUI, building based on jupyter-mcp-server. Thanks to the contributor @ChengJiale150.

• [2025.10.28]: We welcome all contributions, including improving the DeepAnalyze and sharing use cases (see CONTRIBUTION.md). All merged PRs will be listed as contributors.

• [2025.10.27]: DeepAnalyze has attracted widespread attention, gaining 1K+ GitHub stars and 200K+ Twitter views within a week.

• [2025.10.21]: DeepAnalyze's paper, code, model, training data are released!

Upload the data, DeepAnalyze can perform data-oriented deep research 🔍 and any data-centric tasks 🛠

• Clone this repo and download DeepAnalyze-8B.
• Deploy DeepAnalyze-8B via vllm: vllm serve DeepAnalyze-8B
• Run these scripts to launch the API and interface, and then interact through the browser (http://localhost:4000):

  cd demo/chat/frontend
  npm install
  cd ..
  bash start.sh
  
  # stop the api and interface
  bash stop.sh
  
  # stop vllm if needed

• If you want to deploy under a specific IP, please replace localhost with your IP address in ./demo/chat/backend.py and ./demo/chat/frontend/lib/config.ts

Upload the data, DeepAnalyze can perform data-oriented deep research 🔍 and any data-centric tasks 🛠

• A more streamlined UI

• Added support for HeyWhale API keys

• Added support for a Docker-based sandbox code execution environment.

• The usage method is the same as WebUI.

  cd demo/chat_v2/frontend
  npm install
  cd ..
  cp .env.example .env
  bash start.sh
  # stop the api and interface
  bash stop.sh
  
  # stop vllm if needed

Familiar with Jupyter Notebook? Try DeepAnalyze through the JupyterUI!

• This Demo runs Jupyter Lab as frontend, creating a new notebook, converting <Analyze|Understand|Answer> to Markdown cells, converting <Code> to Code cells and executing them as <Execute>.
• Go to demo/jupyter to see more and try!
• 👏Thanks a lot to the contributor @ChengJiale150.

Try DeepAnalyze through the command-line interface

• Deploy DeepAnalyze-8B via vllm: vllm serve DeepAnalyze-8B

• Start the API server and launch the CLI interface:

  cd API
  python start_server.py  # In one terminal
  
  cd demo/cli
  python api_cli.py       # In another terminal (English)
  # or
  python api_cli_ZH.py    # In another terminal (Chinese)

• The CLI provides a Rich-based beautiful interface with file upload support and real-time streaming responses.

• Supports both English and Chinese interfaces .

API keys are now available!

To request your key, please fill out one of the following application forms:

• Primary Form (Google)
• Alternative Form (Feishu)

📚 For comprehensive usage instructions, please refer to the API guide:

• Documentation
• Feishu Wiki

Download model in RUC-DataLab/DeepAnalyze-8B · Hugging Face or DeepAnalyze-8B · ModelScope

To obtain the quantized model, you can use ./quantize.py .

python -m vllm.entrypoints.openai.api_server \
  --model <model_path> \
  --served-model-name DeepAnalyze-8B \
  --max-model-len <select_from_table_above> \
  --gpu-memory-utilization 0.95 \
  --port 8000 \
  <add_fp8_if_required> \
  --trust-remote-code

Scenario 1: 16GB GPU Memory Users (Recommended: 4-bit Quantized Version)

python -m vllm.entrypoints.openai.api_server \
  --model /path/to/deepanalyze/4bit \
  --served-model-name DeepAnalyze-8B \
  --max-model-len 49152 \
  --gpu-memory-utilization 0.95 \
  --port 8000 \
  --kv-cache-dtype fp8 \
  --trust-remote-code

Scenario 2: 24GB GPU Memory Users (For Maximum Context Length)

python -m vllm.entrypoints.openai.api_server \
  --model /path/to/deepanalyze/4bit \
  --served-model-name DeepAnalyze-8B \
  --max-model-len 131072 \
  --gpu-memory-utilization 0.95 \
  --port 8000 \
  --kv-cache-dtype fp8 \
  --trust-remote-code

Scenario 3: 80GB GPU Memory Users (Best Performance)

python -m vllm.entrypoints.openai.api_server \
  --model /path/to/original/model \
  --served-model-name DeepAnalyze-8B \
  --max-model-len 131072 \
  --gpu-memory-utilization 0.95 \
  --port 8000 \
  --trust-remote-code

• Limited Memory (<24GB): Use 4-bit Quantized Version + FP8 KV Cache
• Balanced Configuration (24-40GB): Choose model type based on requirements
• Sufficient Memory (≥ 0GB): Use Original Model for best precision

After launching, the API service can be accessed via http://localhost:8000/v1/completions.

• Install packages: torch, transformers, vllm>=0.8.5

  conda create -n deepanalyze python=3.12 -y
  conda activate deepanalyze
  pip install -r requirements.txt
  
  # For training
  (cd ./deepanalyze/ms-swift/ && pip install -e .)
  (cd ./deepanalyze/SkyRL/ && pip install -e .)

• requirements.txt lists the minimal dependencies required for DeepAnalyze inference. For training, please refer to ./deepanalyze/ms-swift/requirements.txt and ./deepanalyze/SkyRL/pyproject.toml
• We recommend separating the inference and training environments to avoid dependency conflicts.

• Deploy DeepAnalyze-8B via vllm: vllm serve DeepAnalyze-8B

• Run these scripts for any data science tasks:

  • You can specify any data science tasks, including specific data tasks and open-ended data research.
  • You can specify any number of data sources, and DeepAnalyze will automatically explore them.
  • You can specify any type of data sources, e.g., structured data (Databases, CSV, Excel), semi-structured data (JSON, XML, YAML), and unstructured data (TXT, Markdown)

  from deepanalyze import DeepAnalyzeVLLM
  
  prompt = """# Instruction
  Generate a data science report.
  
  # Data
  File 1: {"name": "bool.xlsx", "size": "4.8KB"}
  File 2: {"name": "person.csv", "size": "10.6KB"}
  File 3: {"name": "disabled.xlsx", "size": "5.6KB"}
  File 4: {"name": "enlist.csv", "size": "6.7KB"}
  File 5: {"name": "filed_for_bankrupcy.csv", "size": "1.0KB"}
  File 6: {"name": "longest_absense_from_school.xlsx", "size": "16.0KB"}
  File 7: {"name": "male.xlsx", "size": "8.8KB"}
  File 8: {"name": "no_payment_due.xlsx", "size": "15.6KB"}
  File 9: {"name": "unemployed.xlsx", "size": "5.6KB"}
  File 10: {"name": "enrolled.csv", "size": "20.4KB"}"""
  
  workspace = "/home/u2023000922/zhangshaolei/deepanalyze_public/DeepAnalyze/example/analysis_on_student_loan/"
  
  deepanalyze = DeepAnalyzeVLLM(
      "/fs/fast/u2023000922/zhangshaolei/checkpoints/deepanalyze-8b/"
  )
  answer = deepanalyze.generate(prompt, workspace=workspace)
  print(answer["reasoning"])

  You shoud get a deep research report, which can be rendered as a PDF.:

  # Comprehensive Analysis of Student Enrollment Patterns and Institutional Transfers
  
  ## Introduction and Research Context
  
  The analysis of student enrollment patterns represents a critical area of educational research with significant implications for institutional planning, resource allocation, and student support services. This comprehensive study examines a comprehensive dataset encompassing 1,194 enrollment records across six educational institutions, merged with supplementary demographic, financial, and employment status data. The research employs advanced analytical techniques including network analysis, predictive modeling, and temporal pattern recognition to uncover both macro-level institutional trends and micro-level student mobility patterns. The dataset's longitudinal nature, spanning fifteen months of enrollment records, provides unique insights into the complex dynamics of student pathways through higher education systems.
  
  Our methodological approach combines quantitative analysis of enrollment durations, transfer probabilities, and financial indicators with qualitative ...
  
  The research contributes to the growing body of literature on student mobility by providing empirical evidence of institutional transfer networks and their relationship to student outcomes...
  .....
  

  

  For more examples and task completion details, please refer to DeepAnalyze's homepage.

• You can build an OpenAI-Style API, using this script (note to change MODEL_PATH = "DeepAnalyze-8B" in API/config.py to your vllm model name):

  python API/start_server.py
  

• API usage :

  FILE_RESPONSE=$(curl -s -X POST "http://localhost:8200/v1/files" \
      -F "file=@data.csv" \
      -F "purpose=file-extract")
  
  FILE_ID=$(echo $FILE_RESPONSE | jq -r '.id')
  
  curl -X POST http://localhost:8200/v1/chat/completions \
       -H "Content-Type: application/json" \
       -d "{
          \"model\": \"DeepAnalyze-8B\",
          \"messages\": [
            {
              \"role\": \"user\",
              \"content\": \"Generate a data science report.\",
              \"file_ids\": [\"$FILE_ID\"]
            }
          ]
        }"
  # wait for a while
  

• Refer to API/README.md for details.

• Download DeepSeek-R1-0528-Qwen3-8B. Or you can directly finetune based on DeepAnalyze-8B.

  • If you use DeepSeek-R1-0528-Qwen3-8B as the base model, you should add the special tokens, using:

    MODEL_PATH=path_to_DeepSeek-R1-0528-Qwen3-8B
    SAVE_PATH=path_to_save_DeepSeek-R1-0528-Qwen3-8B-addvocab
    
    python deepanalyze/add_vocab.py \
      --model_path "$MODEL_PATH" \
      --save_path "$SAVE_PATH" \
      --add_tags

• Download training data DataScience-Instruct-500K.

  • unzip DataScience-Instruct-500K/RL/data.zip

• Single-ability Fine-tuning: ./scripts/single.sh
• Multi-ability Agentic Training (cold start): ./scripts/multi_coldstart.sh
• Multi-ability Agentic Training (RL): ./scripts/multi_rl.sh

• We have unified the evaluation of most existing data science benchmarks using vLLM (with more being continuously added...). You can directly follow the introduction in ./playground to quickly evaluate DeepAnalyze or your own agent.

We welcome all forms of contributions, and merged PRs will be listed as contributors.

• We welcome all forms of contributions on DeepAnalyze's code, model and UI, such as Docker packaging, DeepAnalyze model conversion and quantization, and submitting DeepAnalyze workflows based on closed-source LLMs.
• You can submit a pull request directly.
• Please refer to the Developer Guides for contribution guidelines.

• We also especially encourage you to share your use cases and feedback when using DeepAnalyze; these are extremely valuable for helping us improve DeepAnalyze.
• You can place your use cases in a new folder under .example/. We recommend following the folder structure of .example/analysis_on_student_loan/, which includes three parts:
  • data/: stores the uploaded files
  • prompt.txt: input instructions
  • README.md: documentation. We suggest including the input, DeepAnalyze's output, outputs from other closed-source LLMs (optional), and your evaluation/comments of the case.
• DeepAnalyze only has 8B parameters, so we also welcome examples where DeepAnalyze performs slightly worse than the closed-source LLMs — this will help us improve DeepAnalyze.

• Training Frameworks: ms-swift, SkyRL

• Sources of Training Data: Reasoning-Table, Spider, BIRD, DABStep

• API Key & Related Services: HeyWhale Community .

  HeyWhale Community (www.heywhale.com) is a world-leading Chinese hands-on AI community. By providing massive data resources, practical cases, learning materials, and a wide range of AI training activities, it brings together nearly one million AI practitioners and enthusiasts to share insights, exchange ideas, collaborate, and rapidly advance their skills through practice.

If this repository is useful for you, please cite as:

@misc{deepanalyze,
      title={DeepAnalyze: Agentic Large Language Models for Autonomous Data Science},
      author={Shaolei Zhang and Ju Fan and Meihao Fan and Guoliang Li and Xiaoyong Du},
      year={2025},
      eprint={2510.16872},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2510.16872},
}

If you have any questions, please feel free to submit an issue or contact zhangshaolei98@ruc.edu.cn.

Welcome to join the Data+AI Enterprise WeChat group, chat and share ideas with other users.

If you like DeepAnalyze, give it a GitHub Star ⭐