ruc-datalab/DeepAnalyze

DeepAnalyze – an autonomous data‑science agent

What it is – DeepAnalyze is an open‑source, agentic large language model (LLM) built to run the whole data‑science workflow without human hand‑holding. The model (DeepAnalyze‑8B) and its training data (a 500 K instruction set for data‑science tasks) are released on Hugging Face, and the repository ships code for inference, a web UI, a Jupyter UI, a CLI, and an OpenAI‑compatible API.

Key capabilities

• End‑to‑end pipeline – From loading raw tables (CSV, Excel, databases, JSON, etc.) through cleaning, analysis, modelling, visualisation and report generation.
• Open‑ended research – The agent can explore heterogeneous data sources, write and run code, and produce analyst‑grade research reports.
• Multiple front‑ends – A browser‑based WebUI (v1 and a newer v2 with Docker sandboxing), a JupyterLab UI that turns special tags into notebook cells, and a terminal‑based CLI.
• Deployable via vLLM – The model can be served with the vLLM inference engine (vllm serve DeepAnalyze-8B) and accessed through an OpenAI‑style endpoint.
• Extensible – All components (model, data, demo) are open‑source, and the repo includes scripts for quantising the model (4‑bit/8‑bit) and for training‑related packages.

How to get started

1. Download the model from Hugging Face (or ModelScope).
2. Set up an environment (conda create -n deepanalyze python=3.12 then pip install -r requirements.txt).
3. Serve the model with vLLM, choosing a memory‑appropriate configuration (e.g., 4‑bit quantised + FP8 KV cache for 16 GB GPUs).
4. Run a demo – start the API (python API/start_server.py) and launch one of the front‑ends (WebUI, JupyterUI, or CLI) following the short scripts in demo/.
5. Use the API – send file uploads and chat‑completion requests just like the OpenAI API; an API‑key request form is provided.

Why it matters – DeepAnalyze aims to be the first LLM that can autonomously perform data‑science tasks that normally require a human analyst, lowering the barrier for non‑experts to extract insights from messy data. Its open‑source nature lets researchers experiment with agentic LLMs, improve the underlying model, or build custom data‑analysis assistants.

---

Quick reference

• Model: RUC-DataLab/DeepAnalyze-8B (8‑billion parameters) – available quantised.
• Training data: DataScience-Instruct-500K (instruction‑following dataset for data‑science).
• Demo links: WebUI videos, JupyterUI screenshots, CLI demo GIFs in the README.
• Related projects: EvoOntology, SkillAdam, DeepPrep, CoDA‑Bench (all from the same lab, extending the ecosystem).

Feel free to clone the repo, spin up the model, and try the web UI to see an LLM generate a full data‑science report from a set of CSV/Excel files.