// the find
ruc-datalab/DeepAnalyze
DeepAnalyze is the first agentic LLM for autonomous data science. 🎈你的AI数据分析师，自动分析大量数据，一键生成专业分析报告！
DeepAnalyze is a fine-tuned 8B model (based on DeepSeek-R1/Qwen3) trained specifically for autonomous data science tasks — it can take a pile of CSV/Excel/JSON files and produce a structured analysis report without you writing any code. It's a research project from Renmin University with a proper arXiv paper, 500K training examples released publicly, and working demos across WebUI, Jupyter, and CLI interfaces. The target audience is data scientists who want to automate exploratory analysis and report generation, and ML researchers who want to study or extend data-science-specific agent training.
The training pipeline is fully open: model weights, 500K-sample instruction dataset, and the RL training scripts (via ms-swift + SkyRL) are all public, so you can actually reproduce or extend it rather than just consume it. The curriculum-based training approach — single-ability SFT followed by multi-ability cold start and then RL — is methodologically sound and well-documented in the paper. The model runs on consumer hardware (16GB GPU with 4-bit quantization), which puts it in reach for individual researchers without a cluster. The Jupyter integration is a genuine usability win: it maps the model's reasoning steps to notebook cells so you can inspect and modify what it did rather than treating the output as a black box.
Code execution is unsandboxed by default — the README shows a Docker option in v2 but the base setup runs generated Python directly on the host, which is a real problem if you point it at untrusted data or let anyone else use the deployment. The API key story is a mess: there's an apply-via-Google-Form flow for a hosted API, an OpenAI-compatible local API you run yourself, and a HeyWhale third-party integration, and the README doesn't clearly distinguish which path you're on at any given moment. The 8B parameter ceiling is going to hurt on complex multi-table analyses — the sample report in the README sounds plausible but the model will hallucinate statistical claims on datasets it can't fully fit in context, and there's no built-in verification step to catch that. Documentation is split across the README, a Feishu wiki, and an external developer guide URL that may or may not stay up, which makes it harder to trust long-term.